Thursday's tape prints the first Haiku-class frontier model at ten-cent-Mtok on 1M context on the same 24 hours the Anthropic coding CLI makes it the default Haiku on the API and the OpenAI coding CLI makes GPT-6.1 Sol the default model across its bundled plus Amazon Bedrock catalogs. On the model-release tape, Anthropic on Wed Oct 7 ships Claude Haiku 5.5 (model id claude-haiku-5-5) at $0.10 / $0.50 per Mtok input/output under 100k and $0.50 / $2.50 above 100k, with cache reads at $0.01 / $0.05 — about 75% cheaper on average than Haiku 4.5 (90% cheaper under 100k, 50% cheaper above), the first Haiku-class model with an adjustable effort setting (Low, Med, High, Xhigh, Max), on an updated tokenizer; available same-day on AWS, Google Cloud and Microsoft Azure; benchmarks against Haiku 4.5 / GPT-6 Luna / Sonnet 5.5 land at 72.4% / 15.7% / 48.9% / 83.9% on OSWorld 2.1 offline, 39.2% / 0.0% / 16.4% / 70.6% on Terminal-Bench 4.0, and 1620 / 735 / 1437 / 1840 Elo on GDPval-AA v2.1; Asana reports over 30% lower task-completion latency and up to 2.5x faster per-agent-turn inference, Box reports +11 points vs Haiku 4.5 at about half the latency. On the coding-agent runtime tape, Anthropic on Wed Oct 7 18:10 ships Claude Code v2.1.293 which adds Claude Haiku 5.5 (claude-haiku-5-5) as the default Haiku model on the Anthropic API — 1M context, $0.10 / $0.50 per Mtok ($0.50 / $2.50 for prompts over 100K), plus agentType on the subagentStatusLine payload for scripts to tell custom subagent types apart, isDeferred on $.tool.register for mods (false lists the tool's schema in the prompt from the start instead of behind tool search), and a fix for an HTTP MCP memory leak that kept every request a connection had sent until it closed; OpenAI on Wed Oct 7 15:58 ships Codex 0.161.0 — GPT-6.1 Sol is now the default model in the bundled and Amazon Bedrock catalogs, Amazon Bedrock supports multi-agent V2 and Ultra reasoning, /mcp login <name> for MCP server sign-in from an active terminal session, a voice device picker, Daybreak opt-in via --enable cli_daybreak, codex exec --cyber-access-program for per-turn Cyber routing, filesystem escalation can now grant broader write access, and explicit launch permissions survive terminal reconnects and new sessions; and OpenAI continues its daily 0.162.0 alpha cadence — alpha.13 through alpha.18 across Oct 4–7, latest Wed Oct 7 02:04. On the framework reliability tape, Mastra @mastra/core@1.75.0 on Wed Oct 7 ships a Span Query API (filters + cursors + previews + model cost via storage.querySpans(), client.querySpans() and POST /api/observability/spans/query), aggregateTraces() with token and cost measures (ClickHouse / DuckDB / Postgres), self-embedding vector stores for semantic recall (MongoDBVector via autoEmbed), and takes @mastra/connect to 1.0 stable with a Microsoft Teams channel, a single encrypted Discord bot-token credential and connect tool calls recorded as trace spans; Pydantic AI 2.53.0 on Thu Oct 1 patches a HIGH-severity advisory (GHSA-6fqq-452j-qhrp) in ConcurrencyLimitedModel / limit_model_concurrency where streamed requests could keep their concurrency slot after an early exit or after fully consuming stream_text() — “repeated streams could then block every request sharing the limiter”, plus adds ToolCallJudge for assessing tool calls before execution and managed subagents in CLAI2; and Pydantic AI 2.54.0 on Fri Oct 2 adds a double-Esc conversation rewind menu to clai2, Gemini 3 web search sources, re-raises model errors from Temporal activities with their original type in workflow code, and preserves text and reasoning part boundaries in OpenAIChatModel streams. On the eval + orchestration tape, CrewAI 1.15.24 on Wed Oct 7 adds crewai eval to print a markdown brief when run by an agent, Oracle Integrations, an experimental job lifecycle + runner, moves message summarization into SummarizeMessages and centralises and refreshes context windows, while addressing multiple dependency advisories. On the LangChain plumbing tape, four subpackage releases land across Oct 2–7: langchain-core 1.6.7 (Oct 6) includes openai redacted_content in v1 output for bedrock converse, langchain-huggingface 1.2.3 (Oct 7) swaps to a supported Scaleway model in the streaming test, langchain-fireworks 1.7.1 (Oct 7) preserves reasoning in streaming and tool loops, and langchain-text-splitters 1.1.3 (Oct 2) hardens RecursiveJsonSplitter input handling; in parallel langgraph-cli 0.4.33 on Wed Oct 7 adds --image-uri to deploy an already-pushed image and a deploy listeners list command. Throughline: Thu Oct 8 is the day the Haiku-class frontier moves to ten-cent-Mtok with an effort dial and 1M context, the Anthropic coding CLI makes it the default Haiku on the API inside the same 24 hours, the OpenAI coding CLI makes GPT-6.1 Sol the default across bundled plus Amazon Bedrock, the TypeScript framework ships a span-query + trace-cost + stable connect-1.0 observability surface, the Python framework patches a HIGH-severity concurrency-limiter advisory before the week closes, and the eval + orchestration layer makes markdown-brief evals a first-class crewai command.
Thu Oct 8 is the day the Haiku-class frontier moves to ten-cent-Mtok with an effort dial and 1M context, the Anthropic coding CLI makes it the default Haiku on the API inside the same 24 hours, and the OpenAI coding CLI makes GPT-6.1 Sol the default model across its bundled plus Amazon Bedrock catalogs. On the model-release tape, Anthropic on Wed Oct 7 ships Claude Haiku 5.5 (model id claude-haiku-5-5) at $0.10 input / $0.50 output per Mtok under 100k and $0.50 / $2.50 above 100k, with cache reads at $0.01 / $0.05 — about 75% cheaper on average than Haiku 4.5 (90% cheaper under 100k, 50% cheaper above; roughly 90% of Haiku 4.5 requests fell in the first group), on an updated tokenizer similar to Sonnet 5.5 and Opus 5.5, and the first Haiku-class model with an adjustable effort setting (Low, Med, High, Xhigh, Max); available same-day on AWS, Google Cloud and Microsoft Azure; benchmarks against Haiku 4.5 / GPT-6 Luna / Sonnet 5.5 land at 72.4% / 15.7% / 48.9% / 83.9% on OSWorld 2.1 offline, 39.2% / 0.0% / 16.4% / 70.6% on Terminal-Bench 4.0, 1620 / 735 / 1437 / 1840 Elo on GDPval-AA v2.1, 1578 / 614 / 1336 / 1824 on AA-Briefcase v1.1, and 45.9% / 10.2% / — / 56.9% on Humanity's Last Exam no-tools; customer reports: Asana >30% lower task-completion latency and up to 2.5x faster inference per agent turn, HubSpot 92.8% CRM-suite average over three runs, AlphaSense 0.84 vs 0.76 for Haiku 4.5 across 400 queries, Box +11 points vs Haiku 4.5 at about half the latency. On the coding-agent runtime new-defaults tape, Anthropic on Wed Oct 7 at 18:10 ships Claude Code v2.1.293 which “adds Claude Haiku 5.5 (claude-haiku-5-5) as the default Haiku model on the Anthropic API — 1M context, $0.10/$0.50 per Mtok ($0.50/$2.50 for prompts over 100K)”, plus agentType on the subagentStatusLine payload, isDeferred on $.tool.register (false lists the tool's schema in the prompt from the start instead of behind tool search), and a fix for an HTTP MCP memory leak that “kept every request a connection had sent until it closed”; and OpenAI on Wed Oct 7 at 15:58 ships Codex 0.161.0 — “GPT-6.1 Sol is now the default model in the bundled and Amazon Bedrock catalogs”, “Amazon Bedrock supports multi-agent V2 and Ultra reasoning on compatible models”, “Sign in to MCP servers from an active terminal session with /mcp login <name>”, a voice device picker, “Daybreak is opt-in through --enable cli_daybreak or features.cli_daybreak=true”, “Select a Cyber access program per turn with codex exec --cyber-access-program”, “Approved filesystem escalation can now grant broader write access”, and “Explicit launch permissions survive terminal reconnects and new sessions”; and OpenAI continues its daily 0.162.0 alpha cadence — alpha.13 through alpha.18 across Oct 4–7, latest Wed Oct 7 at 02:04. On the framework reliability tape, Mastra @mastra/core@1.75.0 on Wed Oct 7 ships a Span Query API (filters + cursors + previews + model cost via storage.querySpans(), client.querySpans() and POST /api/observability/spans/query), “aggregateTraces() now supports token and cost measures” (ClickHouse / DuckDB / Postgres), self-embedding vector stores for semantic recall (MongoDBVector via autoEmbed), and takes @mastra/connect to 1.0 stable with a Microsoft Teams channel, a single encrypted Discord bot-token credential, and connect tool calls recorded as trace spans; breaking: session.model.switch signature changes to switch(modelId, options?), @mastra/connect integrations→providers, Slack channel id slack→slack-channels; Pydantic AI 2.53.0 on Thu Oct 1 patches a HIGH-severity advisory (GHSA-6fqq-452j-qhrp) in ConcurrencyLimitedModel / limit_model_concurrency where streamed requests could keep their concurrency slot after an early exit or after fully consuming stream_text() — “repeated streams could then block every request sharing the limiter”, plus adds ToolCallJudge to assess tool calls before execution and managed subagents in CLAI2; and Pydantic AI 2.54.0 on Fri Oct 2 adds a double-Esc conversation rewind menu to clai2, builds “the same ModelResponse from streamed and complete GoogleModel responses” with Gemini 3 web search sources, re-raises model errors from Temporal activities with their original type in workflow code, and preserves text and reasoning part boundaries in OpenAIChatModel streams. On the eval + orchestration tape, CrewAI 1.15.24 on Wed Oct 7 adds “crewai eval to print a markdown brief when run by an agent”, Oracle Integrations, an experimental job lifecycle + runner, moves message summarization into SummarizeMessages and centralises and refreshes context windows, while addressing multiple dependency advisories. On the LangChain plumbing tape, four subpackage releases land across Oct 2–7: langchain-core 1.6.7 (Oct 6) “includes openai redacted_content in v1 output for bedrock converse”, langchain-huggingface 1.2.3 (Oct 7) hides the HF Hub API token from repr and swaps to a supported Scaleway model in the streaming test, langchain-fireworks 1.7.1 (Oct 7) “preserves reasoning in streaming and tool loops”, and langchain-text-splitters 1.1.3 (Oct 2) hardens RecursiveJsonSplitter input handling; in parallel langgraph-cli 0.4.33 on Wed Oct 7 adds --image-uri to deploy an already-pushed image and a deploy listeners list command. Throughline: Thu Oct 8 is the day the Haiku-class frontier moves to ten-cent-Mtok with an effort dial and 1M context, the Anthropic coding CLI makes Haiku 5.5 the default Haiku on the API inside the same 24 hours, the OpenAI coding CLI makes GPT-6.1 Sol the default across bundled plus Amazon Bedrock, the TypeScript framework ships a Span-Query + trace-cost + stable-connect-1.0 observability surface, the Python framework patches a HIGH-severity concurrency-limiter advisory before the week closes, and the eval + orchestration layer makes markdown-brief evals a first-class crewai command.
Model-release tape — Anthropic ships Claude Haiku 5.5 (claude-haiku-5-5) on Wed Oct 7 at $0.10/$0.50 per Mtok input/output under 100k and $0.50/$2.50 above 100k with cache reads at $0.01/$0.05, roughly 75% cheaper on average than Haiku 4.5 (90% cheaper under 100k, 50% cheaper above), the first Haiku-class model with an adjustable effort setting (Low/Med/High/Xhigh/Max) on an updated tokenizer, available same-day on AWS + Google Cloud + Microsoft Azure, and benchmarked at 72.4% OSWorld 2.1 offline + 39.2% Terminal-Bench 4.0 + 1620 Elo GDPval-AA v2.1
Anthropic on Wed Oct 7 ships Claude Haiku 5.5 — model id claude-haiku-5-5 — at $0.10 input / $0.50 output per Mtok for prompts up to 100k, $0.50 / $2.50 above 100k, cache writes $0.125 / $0.625, cache reads $0.01 / $0.05; pitched as “about 75% cheaper on average than Haiku 4.5” — 90% cheaper for requests up to 100k tokens and 50% cheaper above, with the page noting roughly 90% of Haiku 4.5 requests fell in the first group — and the first Haiku-class model with an adjustable effort setting (Low, Med, High, Xhigh, Max), on an updated tokenizer similar to Sonnet 5.5 and Opus 5.5 that uses slightly more tokens per task; the page frames Haiku 5.5 for high-volume, cost-sensitive work (summaries, compaction, database queries, classification), as a subagent on coding work paired with Opus 5.5 or Sonnet 5.5, and for speed-sensitive uses such as live customer support and browser use, while pointing to Sonnet 5.5 and Opus 5.5 for complex agentic coding; available same-day on AWS, Google Cloud and Microsoft Azure; benchmarks against Haiku 4.5 / GPT-6 Luna / Sonnet 5.5 land at 72.4% / 15.7% / 48.9% / 83.9% on OSWorld 2.1 offline subset, 39.2% / 0.0% / 16.4% / 70.6% on Terminal-Bench 4.0, 1620 / 735 / 1437 / 1840 Elo on GDPval-AA v2.1, 1578 / 614 / 1336 / 1824 on AA-Briefcase v1.1, 45.9% / 10.2% / — / 56.9% on Humanity's Last Exam no-tools, 57.4% / 18.7% / — / 64.5% with tools, 46.4% / 6.4% / 29.1% / 61.6% on Chartography no-tools; customer-reported: Asana >30% lower task-completion latency and up to 2.5x faster inference per agent turn, HubSpot 92.8% average over three runs on its CRM suite, AlphaSense 0.84 vs 0.76 for Haiku 4.5 across 400 queries, Box +11 points vs Haiku 4.5 at about half the latency; the operative signal that the honest 2026 Haiku-class question has moved from “how fast and cheap can the small-tier model be” to “does the Haiku-class model land at $0.10/$0.50 per Mtok under 100k with an adjustable effort dial, hit 72.4% OSWorld 2.1 offline and 39.2% Terminal-Bench 4.0 against a Haiku 4.5 baseline of 15.7% / 0.0%, and GA same-day on all three major clouds”
Wed Oct 7 2026 · Lab: Anthropic · Model: Claude Haiku 5.5 · Model id: claude-haiku-5-5 · Pricing (under 100k): $0.10 input / $0.50 output per Mtok · Pricing (over 100k): $0.50 / $2.50 per Mtok · Cache writes: $0.125 / $0.625 · Cache reads: $0.01 / $0.05 · Cost shift vs Haiku 4.5: ~75% cheaper on average (90% cheaper under 100k, 50% above) · Tokenizer: updated, similar to Sonnet 5.5 / Opus 5.5 (slightly more tokens per task) · First-of-its-kind: first Haiku-class model with adjustable effort setting (Low/Med/High/Xhigh/Max) · Clouds (same-day): AWS + Google Cloud + Microsoft Azure · OSWorld 2.1 offline: 72.4% (vs Haiku 4.5 15.7% / GPT-6 Luna 48.9% / Sonnet 5.5 83.9%) · Terminal-Bench 4.0: 39.2% (vs 0.0% / 16.4% / 70.6%) · GDPval-AA v2.1 Elo: 1620 (vs 735 / 1437 / 1840) · AA-Briefcase v1.1: 1578 (vs 614 / 1336 / 1824) · HLE no-tools: 45.9% / HLE with tools: 57.4% · Customer reports: Asana >30% latency drop + 2.5x per-turn speed; HubSpot 92.8% CRM; AlphaSense 0.84 vs 0.76; Box +11 at ~half latency · Not pitched for: complex agentic coding (defers to Sonnet 5.5 / Opus 5.5) · Coverage: Anthropic newsTwo reads. (1) A Haiku-class frontier model at $0.10 / $0.50 per Mtok under 100k, with an adjustable effort setting across five levels and same-day GA on AWS + Google Cloud + Microsoft Azure, is the operative signal that the honest 2026 small-tier counter-position has moved from “Haiku-class = fast + cheap, no reasoning dial” to “Haiku-class = $0.10/Mtok with Low→Max effort control, benched against Sonnet 5.5 and GPT-6 Luna, GA on every hyperscaler the same day it posts”. The 90%-cheaper-under-100k tell is the operative pricing-signal — moving the under-100k input price from $1.00/Mtok (Haiku 4.5) to $0.10/Mtok (Haiku 5.5) while the under-100k output drops from $5.00 to $0.50 is the posture a lab takes when it has decided “high-volume, cost-sensitive work” (summaries, compaction, DB queries, classification) and “subagent in a coding pipeline” are the primary SKUs, not “chat completion for a single-user app”. (2) The 72.4%-OSWorld-2.1-plus-39.2%-Terminal-Bench-4.0 tell is the operative capability-signal — a Haiku-class model going from Haiku 4.5's 15.7% / 0.0% on those same benches to 72.4% / 39.2% inside a single generation, while sitting at 83.9% / 70.6% on Sonnet 5.5, is the shape a lab takes when it has decided the Haiku-class ceiling must leap rather than drift. Landing on the same 24 hours as Claude Code 2.1.293 making claude-haiku-5-5 the default Haiku on the Anthropic API with 1M context (item 02) and Codex 0.161 making GPT-6.1 Sol the default on bundled + Amazon Bedrock (item 03), Haiku 5.5 becomes the reference “the Haiku-class frontier lands at $0.10/Mtok on an effort dial with same-day three-cloud GA on the same day the coding CLIs change their defaults” primitive every subsequent GPT-6 Luna, Gemini 3 Flash, Grok 4 Mini and Mistral Small response now has to price against.
Coding-agent runtime new-defaults tape — Claude Code 2.1.293 lands at 18:10 UTC and “adds Claude Haiku 5.5 (claude-haiku-5-5), now the default Haiku model on the Anthropic API — 1M context, $0.10/$0.50 per Mtok”, plus agentType on subagentStatusLine and isDeferred on $.tool.register; Codex 0.161.0 lands at 15:58 UTC the same day and makes GPT-6.1 Sol the default across its bundled plus Amazon Bedrock catalogs, with Bedrock multi-agent V2 + Ultra reasoning, /mcp login <name>, Daybreak opt-in, codex exec --cyber-access-program for per-turn Cyber routing, filesystem escalation + launch-permission persistence; and OpenAI continues its daily 0.162.0 alpha cadence from alpha.13 to alpha.18 across Oct 4-7
Anthropic on Wed Oct 7 at 18:10 ships Claude Code v2.1.293 — the first Claude Code release to carry the Haiku 5.5 default — whose load-bearing release-note bullet reads “Added Claude Haiku 5.5 (claude-haiku-5-5), now the default Haiku model on the Anthropic API — 1M context, $0.10/$0.50 per Mtok ($0.50/$2.50 for prompts over 100K)”; the same release adds “agentType to the subagentStatusLine payload, so scripts can tell custom subagent types apart” and “isDeferred to $.tool.register for mods: false lists the tool's schema in the prompt from the start instead of behind tool search”, and fixes “a memory leak where an HTTP MCP connection kept every request it had sent until it closed”; per the GitHub anthropics/claude-code release page; the operative signal that the honest 2026 coding-agent-default question has moved from “does the CLI pick a Haiku model” to “does the coding CLI change its default Haiku to claude-haiku-5-5 inside the same 24 hours the model lands on the API, raise the context to 1M, publish the exact $0.10/$0.50 pricing in the release note, and ship an isDeferred-tool-register primitive so mods can gate schemas behind tool search”
Wed Oct 7 2026 18:10 UTC · Vendor: Anthropic · Release: Claude Code v2.1.293 · New default Haiku: claude-haiku-5-5 (1M context) · Pricing published in release note: $0.10/$0.50 per Mtok under 100K, $0.50/$2.50 above · New mod hook fields: agentType on subagentStatusLine + isDeferred on $.tool.register · Notable fix: HTTP MCP memory leak (connection kept every request it sent until close) + claude purge change · Delivery: npm + Plugins framework · Coverage: GitHub anthropics/claude-code releasesTwo reads. (1) A frontier lab making claude-haiku-5-5 the default Haiku on the Anthropic API inside the same 24 hours the model lands, with the exact $0.10/$0.50 pricing printed in the release note and 1M context on by default, is the operative signal that the honest 2026 coding-agent-default counter-position has moved from “the model lands this quarter; the default swaps next release” to “model-release and CLI-default land on the same date”. The isDeferred-on-$.tool.register tell is the operative mod-API signal — letting a mod declare “false lists the tool's schema in the prompt from the start instead of behind tool search” is the posture a vendor takes when it has decided plugin-authored tool-schema placement is a first-class control, not a hardcoded policy. (2) The HTTP-MCP-connection-kept-every-request-until-close tell is the operative reliability-signal — closing a leak that only shows under long-running HTTP MCP sessions with high request counts is the shape a vendor takes when it has decided the HTTP transport is the production MCP surface, not stdio. Landing on the same 24 hours as Haiku 5.5 itself (item 01) and Codex 0.161 defaulting to GPT-6.1 Sol on bundled + Amazon Bedrock (item 03), Claude Code 2.1.293 becomes the reference “the coding CLI swaps its default Haiku on the same day the model ships, raises the context to 1M, and prints the pricing in the release note” primitive every subsequent OpenAI Codex, Cursor, Cline, Aider, Continue, Zed Agent and Gemini CLI print now has to price against.
OpenAI on Wed Oct 7 at 15:58 ships Codex 0.161.0 — the first Codex release of the week to change the default model across both the bundled and Amazon Bedrock catalogs — whose load-bearing release-note bullets read “GPT-6.1 Sol is now the default model in the bundled and Amazon Bedrock catalogs”, “Amazon Bedrock supports multi-agent V2 and Ultra reasoning on compatible models”, “Sign in to MCP servers from an active terminal session with /mcp login <name>”, “Choose your microphone, speaker, and microphone input channels for voice conversations”, “Daybreak is opt-in through --enable cli_daybreak or features.cli_daybreak=true”, “Select a Cyber access program per turn with codex exec --cyber-access-program”, “Approved filesystem escalation can now grant broader write access”, and “Explicit launch permissions survive terminal reconnects and new sessions”; in parallel the vendor continues its daily 0.162.0 alpha cadence — alpha.13 (Oct 4 15:42) through alpha.18 (Oct 7 02:04), with alpha.17.1 at Oct 7 23:56 — while a 0.161.0-alpha.13.1 side-branch lands Oct 6 05:27; per the GitHub openai/codex releases page; the operative signal that the honest 2026 coding-agent-default question has moved from “does the CLI run on a frontier model” to “does the coding CLI change its default model on both bundled and Amazon Bedrock catalogs to GPT-6.1 Sol, add Bedrock multi-agent V2 + Ultra reasoning, publish /mcp login as a terminal-session MCP sign-in flow, Daybreak as an --enable flag, and codex exec --cyber-access-program as a per-turn Cyber routing control, all in one release”
Wed Oct 7 2026 15:58 UTC · Vendor: OpenAI · Release: Codex 0.161.0 (tag rust-v0.161.0) · New default model: GPT-6.1 Sol in bundled + Amazon Bedrock catalogs · Bedrock: multi-agent V2 + Ultra reasoning on compatible models · MCP sign-in: /mcp login <name> from an active terminal session · Voice: microphone + speaker + input-channel picker · Daybreak opt-in: --enable cli_daybreak or features.cli_daybreak=true · Per-turn Cyber routing: codex exec --cyber-access-program · Filesystem: approved escalation can grant broader write access · Launch permissions: survive terminal reconnects + new sessions · Parallel alpha cadence: 0.162.0-alpha.13 (Oct 4 15:42) → alpha.14 (Oct 5 02:05) → alpha.15 (Oct 5 13:10) → alpha.16 (Oct 5 21:37) → alpha.17 (Oct 6 21:58) → alpha.17.1 (Oct 7 23:56) → alpha.18 (Oct 7 02:04); side-branch 0.161.0-alpha.13.1 (Oct 6 05:27) · Coverage: GitHub openai/codex releasesTwo reads. (1) A coding-agent CLI making GPT-6.1 Sol the default across both the bundled and Amazon Bedrock catalogs in one release, while adding Bedrock multi-agent V2 + Ultra reasoning on compatible models and a /mcp login <name> terminal-session MCP sign-in flow, is the operative signal that the honest 2026 coding-agent-default counter-position has moved from “Bedrock is a secondary runtime” to “Bedrock gets a parity default and multi-agent V2 + Ultra reasoning in the same release the bundled catalog gets its new default”. The codex-exec---cyber-access-program-per-turn tell is the operative routing-signal — a per-turn Cyber-access-program flag on codex exec is the posture a vendor takes when it has decided “Cyber-access routing is a per-turn decision”, not a session-level setting, and anchors the safety-ladder-aware coding CLI on per-turn access selection. (2) The Daybreak-opt-in-through---enable-cli_daybreak tell is the operative preview-signal — shipping a feature behind --enable cli_daybreak or features.cli_daybreak=true is a very different posture than “GA on the latest tag”, and anchors the vendor on flag-gated preview flows for a daily-alpha-cadenced CLI. Landing on the same 24 hours as Haiku 5.5 (item 01), Claude Code 2.1.293 (item 02) and the Codex 0.162.0 alpha train (item 04), Codex 0.161.0 becomes the reference “the coding CLI changes its default across bundled + Amazon Bedrock to GPT-6.1 Sol inside the same 24 hours the frontier lab lands a $0.10/Mtok Haiku and the peer CLI makes it its default” primitive every subsequent OpenAI Codex, Cursor, Cline, Aider, Continue, Zed Agent and Gemini CLI print now has to price against.
OpenAI between Sun Oct 4 and Wed Oct 7 ships Codex 0.162.0 alpha.13 → alpha.18 on top of the 0.161.0 stable, running a six-stage daily alpha train that lands alpha.13 Oct 4 at 15:42, alpha.14 Oct 5 at 02:05, alpha.15 Oct 5 at 13:10, alpha.16 Oct 5 at 21:37, alpha.17 Oct 6 at 21:58, alpha.17.1 Oct 7 at 23:56, and alpha.18 Oct 7 at 02:04; alongside the 0.162 train, a 0.161.0-alpha.13.1 side-branch lands Oct 6 at 05:27; per the GitHub openai/codex releases page; the operative signal that the honest 2026 coding-agent-release-cadence question has moved from “does the CLI publish a monthly stable” to “does the CLI publish six 0.162.0 alpha tags and a 0.161.0 side-branch alpha across four days, land its new stable mid-train, and keep the daily alpha cadence running after the stable ships”
Sun Oct 4–Wed Oct 7 2026 · Vendor: OpenAI · Release line: Codex 0.162.0 alpha · Six tags: alpha.13 (Oct 4 15:42) → alpha.14 (Oct 5 02:05) → alpha.15 (Oct 5 13:10) → alpha.16 (Oct 5 21:37) → alpha.17 (Oct 6 21:58) → alpha.17.1 (Oct 7 23:56) → alpha.18 (Oct 7 02:04) · Parallel side-branch: 0.161.0-alpha.13.1 (Oct 6 05:27) · Stable in the middle: 0.161.0 (Oct 7 15:58, GPT-6.1 Sol default) · Prior-week alpha context: 0.162.0-alpha.9→.17 ran Oct 3–6 (covered in prior edition) · Coverage: GitHub openai/codex releasesTwo reads. (1) A coding-agent CLI running a six-stage 0.162.0-alpha train across Oct 4–7 while cutting a new stable mid-train and continuing the alpha cadence after the stable ships, with a 0.161.0-alpha.13.1 side-branch landing on the day of the Codex 0.161 stable, is the operative signal that the honest 2026 coding-agent-release-cadence counter-position has moved from “monthly stables + weekly previews” to “daily alphas on the next minor + maintenance alphas on the current minor, with the stable cut in the middle of the train”. The six-tag-train-across-four-days tell is the operative cadence-signal — publishing alpha.13 through alpha.18 across four days, with alpha.17.1 landing the same day as the 0.161 stable, is the shape a vendor takes when it has decided “daily alpha + stable cut mid-train” is the honest release velocity for a coding-agent runtime that other coding tools, mods and plugins ride. (2) The 0.161.0-alpha.13.1-side-branch-the-day-of-0.161-stable tell is the operative release-shape signal — cutting a maintenance alpha on the previous minor on the same day the next minor's stable ships is the posture a vendor takes when it has decided enterprise users pinned to 0.160.x need a side-branch alpha path even as 0.161 goes stable. Landing on the same 24 hours as Claude Code 2.1.293 (item 02) and the Codex 0.161 stable (item 03), the 0.162.0 alpha train becomes the reference “the coding CLI publishes alpha.13-18 across four days, keeps a 0.161.0-alpha.13.1 side-branch alive on stable day, and keeps the alpha cadence running after the stable ships” primitive every subsequent Cursor, Cline, Aider, Continue, Zed Agent, Gemini CLI and Grok Build CLI print now has to price against.
Framework reliability tape — Mastra @mastra/core@1.75.0 ships a Span Query API (storage.querySpans() + client.querySpans() + POST /api/observability/spans/query) with filters, cursors, previews and cost, aggregateTraces() with token + cost measures on ClickHouse / DuckDB / Postgres, self-embedding vector stores via MongoDBVector autoEmbed, and takes @mastra/connect to 1.0 stable with Microsoft Teams channel + single encrypted Discord bot-token credential + connect tool-calls recorded as trace spans; Pydantic AI 2.53.0 patches a HIGH-severity concurrency-slot-leak advisory (GHSA-6fqq-452j-qhrp) in ConcurrencyLimitedModel / limit_model_concurrency and adds ToolCallJudge + managed subagents; Pydantic AI 2.54.0 adds a double-Esc conversation rewind, Gemini 3 web search sources, Temporal error type re-raise and OpenAI stream part-boundary preservation
Mastra @mastra/core@1.75.0 on Wed Oct 7 ships a Span Query API that exposes per-span filters, cursors, previews and model cost through storage.querySpans() on the core, client.querySpans() on the client, and a new POST /api/observability/spans/query HTTP endpoint; “aggregateTraces() now supports token and cost measures” with token sums and averages + cost measures implemented in the ClickHouse, DuckDB and Postgres observability stores; “Semantic recall can now run against vector stores that generate embeddings themselves”, including MongoDBVector via autoEmbed, so recall works without a client-side embedder; @mastra/connect reaches 1.0 stable with a reworked provider API that adds Microsoft Teams to channels(), moves Discord to a single encrypted API-key bot-token credential, and records connect tool calls as trace spans; AgentController sessions keep one active model per thread across mode switches, and sessions can start without creating a thread via createInitialThread: false; breaking changes land in the same release: session.model.switch signature changes to switch(modelId, options?) (modeId and scope removed; mode switches no longer change the model automatically), @mastra/connect integrations→providers (legacy connect() alias removed), “Discovered MCP tools no longer require approval by default” (requireApproval replaces autoApproveTools), and the Slack channel id changes from slack to slack-channels in channels(); per the GitHub mastra-ai/mastra release page; the operative signal that the honest 2026 agent-framework-observability question has moved from “does the framework emit traces” to “does the framework ship a Span Query API with filters + cursors + previews + model cost, aggregate token and cost measures on ClickHouse + DuckDB + Postgres, let semantic recall run against self-embedding vector stores like MongoDBVector autoEmbed, and reach @mastra/connect 1.0 stable with Microsoft Teams + a single encrypted Discord bot-token + tool-calls as trace spans”
Wed Oct 7 2026 · Framework: Mastra @mastra/core@1.75.0 · Span Query API: filters + cursors + previews + model cost via storage.querySpans() + client.querySpans() + POST /api/observability/spans/query · aggregateTraces(): token sums/averages + cost measures (ClickHouse + DuckDB + Postgres) · Self-embedding semantic recall: MongoDBVector via autoEmbed · @mastra/connect 1.0 stable: reworked provider API + Microsoft Teams channel + single encrypted Discord bot-token credential + connect tool-calls as trace spans · AgentController sessions: one active model per thread across mode switches + createInitialThread:false · Breaking: session.model.switch signature (switch(modelId, options?); modeId + scope removed; mode switches no longer change the model) · Breaking: @mastra/connect integrations → providers; connect() alias removed; requireApproval replaces autoApproveTools (discovered MCP tools no longer require approval by default); Slack channel id slack → slack-channels · Coverage: GitHub mastra-ai/mastra releasesTwo reads. (1) A TypeScript agent framework shipping a Span Query API with filters + cursors + previews + model cost, aggregateTraces() token + cost measures on three stores, self-embedding vector stores via MongoDBVector autoEmbed, and @mastra/connect 1.0 stable with Microsoft Teams + a single encrypted Discord bot-token credential + tool-calls as trace spans, inside one release, is the operative signal that the honest 2026 agent-framework-observability counter-position has moved from “traces come out, you grep them” to “spans are a query-shaped primitive with model cost + token measures per aggregate, and connect tool-calls land in the trace stream alongside agent steps”. The MongoDBVector-autoEmbed tell is the operative recall-signal — letting a vector store generate its own embeddings is the posture a framework takes when it has decided semantic recall must survive without a client-side embedder, which anchors the recall surface on whatever the store already understands. (2) The session.model.switch-signature-and-integrations→providers-rename tell is the operative breaking-change signal — a breaking-signature session API + a subpackage integrations→providers rename + a default-off MCP-tool approval flip + a Slack channel-id rename all in one release is the shape a framework takes when it has decided the provider surface and session API are worth breaking to land the observability + connect-1.0 story cleanly. Landing on the same 24 hours as Haiku 5.5 (item 01), Claude Code 2.1.293 (item 02) and Codex 0.161 (item 03), Mastra 1.75 becomes the reference “the TypeScript framework ships Span-Query + trace-cost + self-embedding recall + @mastra/connect 1.0 stable the same day the Haiku-class frontier lands at $0.10/Mtok and the two coding CLIs change their defaults” primitive every subsequent OpenAI Agents SDK, Pydantic AI, Strands, Letta and Archer release now has to price against.
Pydantic AI 2.53.0 on Thu Oct 1 patches GHSA-6fqq-452j-qhrp, a HIGH-severity advisory in ConcurrencyLimitedModel / limit_model_concurrency: “Streamed requests through ConcurrencyLimitedModel or limit_model_concurrency could keep their concurrency slot after an early exit or after fully consuming stream_text()” — “Repeated streams could then block every request sharing the limiter”; agent-level limits and non-streaming requests were not affected, and the fix is “Patched in 2.53.0; v1 is not affected”; the release additionally raises a UserError when a model wrapper shares a limiter with the agent or an enclosing wrapper, makes ConcurrencyLimiter.acquire() take a slot on every call, and adds ToolCallJudge — “Add ToolCallJudge to assess tool calls before execution”, managed subagents in CLAI2 with Claude and Codex agent definitions in Harness, and built-in CLAI2 plugins for PostHog (“/keys or browser sign-in”), Grain (“keyring-backed Grain sign-in”) and Linear (“settings menu and /keys credentials”); streamed and complete responses were aligned for OpenAIChatModel, OpenRouterModel, OpenAIResponsesModel and XaiModel, and Logfire Temporal spans were made replay-safe; per the GitHub Security Advisory and the pydantic/pydantic-ai release page; the operative signal that the honest 2026 agent-framework-reliability question has moved from “does the limiter block over the limit” to “does the limiter patch a HIGH-severity concurrency-slot-leak on streamed requests, raise UserError when a wrapper shares its limiter with the agent, and still ship ToolCallJudge + managed subagents + three CLAI2 plugins in the same release”
Thu Oct 1 2026 (release stamped 02 Oct 02:52 UTC) · Framework: Pydantic AI v2.53.0 · Advisory: GHSA-6fqq-452j-qhrp (HIGH) · Affected: ConcurrencyLimitedModel + limit_model_concurrency on streamed requests (early exit or full stream_text() consumption kept the slot) · Blast: repeated streams could block every request sharing the limiter · Not affected: agent-level limits + non-streaming requests + v1 · Fix: patched in 2.53.0 · Additional behavior: model wrapper raises UserError when sharing a limiter with the agent or enclosing wrapper; ConcurrencyLimiter.acquire() takes a slot on every call · Added: ToolCallJudge (PR 9041); managed subagents in CLAI2 with Claude + Codex agent definitions in Harness (PR 9573); CLAI2 plugins — PostHog (/keys or browser), Grain (keyring-backed sign-in), Linear (/keys + settings menu); oneOf schema support in TestModel generated data (PR 8783) · Streaming alignment: OpenAIChatModel + OpenRouterModel + OpenAIResponsesModel + XaiModel · Observability: Logfire Temporal spans replay-safe · Coverage: GitHub security advisory + pydantic/pydantic-ai release pageTwo reads. (1) A Python agent framework patching a HIGH-severity concurrency-slot-leak advisory on streamed requests in ConcurrencyLimitedModel / limit_model_concurrency, where “Repeated streams could then block every request sharing the limiter”, is the operative signal that the honest 2026 agent-framework-reliability counter-position has moved from “the limiter blocks over the limit” to “the limiter also releases its slot on an early exit or after stream_text() consumption, raises a UserError when a wrapper shares the agent's limiter, and takes a slot on every acquire”. The streamed-vs-non-streamed-asymmetry tell is the operative concurrency-signal — the advisory affects streamed requests but not non-streaming or agent-level limits, which is the posture a framework takes when it has decided streaming is a first-class async path with its own slot-management invariants, not an extension of the sync path. (2) The ToolCallJudge-plus-managed-subagents-plus-three-CLAI2-plugins-same-release tell is the operative scope-signal — shipping a pre-execution tool-call judge, managed subagents with Claude + Codex agent definitions in Harness, and built-in PostHog + Grain + Linear plugins in the same release as a HIGH-severity security patch is a very different cadence than “security patch, then feature cycle”. Landing in the same week as Pydantic AI 2.54.0 (item 07), Mastra 1.75 (item 05) and the coding-CLI default swaps (items 02–03), Pydantic AI 2.53 becomes the reference “the Python framework ships a HIGH-severity concurrency-slot-leak patch + ToolCallJudge + managed subagents + three CLAI2 plugins the same week the TypeScript peer takes @mastra/connect to 1.0” primitive every subsequent LlamaIndex, OpenAI Agents SDK, Letta, Agno, Strands and DSPy release now has to price against.
Pydantic AI 2.54.0 on Fri Oct 2 moves the Python agent framework forward on four reliability axes: a double-Esc conversation rewind menu in clai2, a schema-handling fix that “Handle(s) draft-7 list-form items in JsonSchemaTransformer, TestModel and Mistral streamed output” while making TestModel accept boolean subschemas and capping prefixItems at maxItems, an ImageGeneration-under-FallbackModel-and-dataclasses.replace fix and a separate ImageGenerationTool-to-local-tool fallback on Gemini text models; a Temporal fix that “Re-raise(s) model errors from Temporal model activities with their original type in workflow code”; a Gemini fix that “Build(s) the same ModelResponse from streamed and complete GoogleModel responses, and add(s) Gemini 3 web search sources”; and an OpenAI fix that “Preserve(s) text and reasoning part boundaries in OpenAIChatModel streams”; per the pydantic/pydantic-ai release page; the operative signal that the honest 2026 agent-framework-correctness question has moved from “does the stream complete” to “does the framework build the same ModelResponse from streamed and complete GoogleModel runs with Gemini 3 web search sources, preserve text + reasoning part boundaries in OpenAIChatModel streams, and re-raise Temporal model errors with their original type into workflow code”
Fri Oct 2 2026 (release stamped 03 Oct 03:21 UTC) · Framework: Pydantic AI v2.54.0 · clai2: double-Esc conversation rewind menu (PR 9688) · Schema: draft-7 list-form items + boolean subschemas + prefixItems capped at maxItems in JsonSchemaTransformer / TestModel / Mistral streamed (PR 9560) · Image: ImageGeneration under FallbackModel + dataclasses.replace (PR 9477); ImageGenerationTool to local tool on Gemini text models (PR 9551) · Temporal: re-raise model errors from Temporal activities with original type in workflow code (PR 9580) · Gemini: same ModelResponse from streamed + complete GoogleModel; adds Gemini 3 web search sources (PR 9594) · OpenAI: preserve text + reasoning part boundaries in OpenAIChatModel streams (PR 8729) · Coverage: pydantic/pydantic-ai release pageTwo reads. (1) A Python agent framework shipping Gemini-3-web-search-sources-plus-same-ModelResponse-from-streamed-and-complete-GoogleModel, text-plus-reasoning-part-boundary preservation in OpenAIChatModel streams, and Temporal-activity-model-error re-raise with original type into workflow code, is the operative signal that the honest 2026 agent-framework-streaming counter-position has moved from “the streamed path returns something close to the complete path” to “the streamed path builds the same ModelResponse as the complete path across GoogleModel, preserves text + reasoning boundaries across OpenAIChatModel, and surfaces Temporal activity errors into workflow code with their original type”. The Temporal-model-error-re-raise-with-original-type tell is the operative durable-execution signal — re-raising model-activity errors with their original type into workflow code is the posture a framework takes when it has decided durable-workflow-visible model errors must preserve their type across the activity boundary, not collapse to a generic activity failure. (2) The double-Esc-conversation-rewind-menu-in-clai2 tell is the operative CLI-UX signal — a CLI-level conversation-rewind flow is the shape a framework takes when it has decided the CLI harness is a first-class product surface, not an afterthought. Landing in the same week as Pydantic AI 2.53's HIGH-severity concurrency patch (item 06), Mastra 1.75 (item 05) and CrewAI 1.15.24 (item 08), Pydantic AI 2.54 becomes the reference “the Python framework aligns streamed and complete Gemini ModelResponses with Gemini 3 web search sources, preserves OpenAIChatModel stream boundaries, and re-raises Temporal model errors with original type the same week the HIGH concurrency advisory is patched” primitive every subsequent LlamaIndex, OpenAI Agents SDK and DSPy release now has to price against.
Eval + orchestration tape — CrewAI 1.15.24 adds “crewai eval to print a markdown brief when run by an agent”, Oracle Integrations, an experimental job lifecycle + runner, moves message summarization into SummarizeMessages, centralises and refreshes context windows, and addresses multiple dependency advisories; v1.15.25 follows the same evening with the snapshot + changelog
CrewAI 1.15.24 on Wed Oct 7 at 17:39 adds “crewai eval to print a markdown brief when run by an agent”, Oracle Integrations, an experimental job lifecycle and runner, moves message summarization into SummarizeMessages and centralises and refreshes context windows, while addressing multiple dependency advisories and updating relevant packages; the v1.15.25 snapshot and changelog follows the same evening at 23:52; per the GitHub crewAIInc/crewAI release page; the operative signal that the honest 2026 agent-orchestration-eval question has moved from “does the framework run an eval harness” to “does the orchestration framework add a crewai-eval markdown-brief primitive that runs when an agent invokes it, Oracle Integrations as a first-class connector, an experimental job lifecycle + runner, and refactor message summarization + context-window handling in the same release”
Wed Oct 7 2026 17:39 UTC · Framework: CrewAI 1.15.24 (snapshot 1.15.25 at 23:52 the same day) · Features: crewai eval markdown brief when run by an agent + Oracle Integrations + experimental job lifecycle + runner · Refactors: message summarization into SummarizeMessages + centralise and refresh context windows · Bug fixes: multiple dependency advisories addressed · Coverage: GitHub crewAIInc/crewAI releasesTwo reads. (1) An orchestration framework shipping “crewai eval to print a markdown brief when run by an agent” as a top-line feature, with an experimental job lifecycle + runner, is the operative signal that the honest 2026 agent-orchestration-eval counter-position has moved from “eval is a side harness you wire up by hand” to “crewai eval is a first-class command the agent itself can call, which prints a markdown brief as its native output”. The message-summarization-into-SummarizeMessages-plus-centralised-context-windows tell is the operative plumbing-signal — moving summarization into a dedicated class and centralising and refreshing context-window handling in the same release is the posture a framework takes when it has decided “context compaction is a core subsystem”, not a thin helper function sprinkled through the loop. (2) The Oracle-Integrations-as-a-named-top-line-feature tell is the operative connector-signal — shipping Oracle Integrations in the same release as crewai-eval and job-lifecycle is the shape a framework takes when it has decided the enterprise Oracle connector surface is worth a first-class slot alongside eval infrastructure. Landing on the same 24 hours as Haiku 5.5 (item 01), Claude Code 2.1.293 (item 02) and Mastra 1.75 (item 05), CrewAI 1.15.24 becomes the reference “the orchestration framework makes crewai-eval a markdown-brief primitive and ships Oracle Integrations the same day the Haiku-class frontier lands at $0.10/Mtok” primitive every subsequent LangGraph, LlamaIndex, Mastra, Pydantic AI and OpenAI Agents SDK eval story now has to price against.
LangChain + LangGraph plumbing tape — four LangChain subpackage releases land Oct 2-7 (core 1.6.7 ships “openai redacted_content in v1 output for bedrock converse”, huggingface 1.2.3 hides the HF Hub token from repr and swaps to a supported Scaleway streaming test, fireworks 1.7.1 “preserves reasoning in streaming and tool loops”, text-splitters 1.1.3 hardens RecursiveJsonSplitter input handling); langgraph-cli 0.4.33 adds --image-uri to deploy an already-pushed image and a deploy listeners list command
LangChain ships four subpackage releases across Oct 2–7: langchain-core 1.6.7 on Tue Oct 6 whose load-bearing fix is “fix(core): include openai redacted_content in v1 output for bedrock converse”; langchain-huggingface 1.2.3 on Wed Oct 7 whose top release note is “fix(huggingface): use a supported Scaleway model in streaming test” (the release also hides the HF Hub API token from repr, removes the retired IPEX backend, and uses torch.accelerator for device-count detection); langchain-fireworks 1.7.1 on Wed Oct 7 whose notable fix is “fix(fireworks): preserve reasoning in streaming and tool loops”; and langchain-text-splitters 1.1.3 on Fri Oct 2 which raises a TypeError for non-dict, non-convertible input to RecursiveJsonSplitter; per the GitHub langchain-ai/langchain release page; the operative signal that the honest 2026 LangChain-subpackage question has moved from “does the subpackage ship with the monorepo” to “does the subpackage train close an openai-redacted-content gap in v1 output for Bedrock Converse, preserve reasoning in Fireworks streaming + tool loops, hide the Hugging Face Hub API token from repr, and raise TypeError for non-dict RecursiveJsonSplitter inputs, all inside a six-day Oct 2-7 window”
Fri Oct 2–Wed Oct 7 2026 · Project: langchain-ai/langchain (subpackage train) · langchain-core 1.6.7 (Oct 6): include openai redacted_content in v1 output for bedrock converse + Python 3.14 inspect.signature hardening · langchain-huggingface 1.2.3 (Oct 7): HF Hub API token hidden from repr + retired IPEX backend removed + torch.accelerator for device-count detection + supported Scaleway model in streaming test · langchain-fireworks 1.7.1 (Oct 7): preserve reasoning content during streaming and tool loops · langchain-text-splitters 1.1.3 (Oct 2): raises TypeError for non-dict, non-convertible input to RecursiveJsonSplitter · Coverage: GitHub langchain-ai/langchain releasesTwo reads. (1) A monorepo running a four-subpackage release train across six days, with each subpackage shipping a narrow, named fix (openai redacted_content in v1 output for Bedrock Converse, reasoning preserved in Fireworks streaming + tool loops, HF Hub token hidden from repr, TypeError on non-dict RecursiveJsonSplitter input), is the operative signal that the honest 2026 LangChain-subpackage counter-position has moved from “the monorepo cuts periodic combined releases” to “each subpackage ships on its own cadence with a single named fix per cut”. The reasoning-preserved-in-streaming-and-tool-loops tell is the operative streaming-signal — the Fireworks provider explicitly “preserving reasoning content during streaming and tool loops” is the posture a subpackage takes when it has decided reasoning content is a first-class part of the stream and must survive the tool-use round-trip, not get lost in the handoff. (2) The HF-Hub-token-hidden-from-repr tell is the operative hygiene-signal — hiding a provider API token from repr is the shape a subpackage takes when it has decided “never let the token print in a debugger or an exception trace” is a correctness invariant, not a lint. Landing in the same week as Mastra 1.75 (item 05), Pydantic AI 2.53-2.54 (items 06–07) and CrewAI 1.15.24 (item 08), the LangChain subpackage train becomes the reference “the monorepo cuts four subpackage releases across six days with a single named fix each, closing an openai-redacted-content gap in Bedrock Converse output and preserving reasoning in Fireworks streaming + tool loops” primitive every subsequent LlamaIndex, Haystack, Semantic Kernel and Agno subpackage release now has to price against.
LangGraph langgraph-cli 0.4.33 on Wed Oct 7 at 13:38 adds “--image-uri to deploy an already-pushed image” and a “deploy listeners list command”; per the GitHub langchain-ai/langgraph release page; the operative signal that the honest 2026 agent-graph-deploy question has moved from “does the CLI build and push an image for me” to “does the CLI accept a pre-pushed --image-uri to deploy an already-built image and ship a deploy-listeners list command for inspecting the deployed listener fleet”
Wed Oct 7 2026 13:38 UTC · Project: langchain-ai/langgraph · Release: langgraph-cli 0.4.33 · New: --image-uri to deploy an already-pushed image · New: deploy listeners list command · Coverage: GitHub langchain-ai/langgraph releasesTwo reads. (1) A graph-framework CLI adding --image-uri to deploy an already-pushed image is the operative signal that the honest 2026 agent-graph-deploy counter-position has moved from “the CLI must build the image for you” to “the CLI accepts a pre-pushed image URI and deploys from it directly, so a separate CI pipeline can own the build”; the deploy-listeners-list tell is the operative introspection-signal — shipping a list command for the deployed listener fleet is the shape a CLI takes when it has decided “the deployment is a managed inventory the operator queries from the terminal”, not a one-shot push-and-pray. (2) The pre-pushed-image-URI tell is the operative CI-boundary signal — letting a CI pipeline own the image build and the CLI own the deploy step is a cleaner separation of concerns than the “CLI does both” frame, and anchors the LangGraph deploy surface on “image URIs are the deploy contract”. Landing in the same week as Mastra 1.75's Span Query API (item 05), the Pydantic AI releases (items 06–07), CrewAI 1.15.24 (item 08) and the LangChain subpackage train (item 09), langgraph-cli 0.4.33 becomes the reference “the graph CLI lets a pre-pushed image URI own the deploy and ships a deploy-listeners list command in the same week the TypeScript peer takes @mastra/connect to 1.0” primitive every subsequent Inngest, Trigger.dev, Hatchet, Restate and Temporal Workflow deploy-UX print now has to price against.