Day 14 — OpenAI tapes out Jalapeño, Codex cuts v0.142.x stable
— Gemini grows native eyes, Anthropic's flagships sit dark on Liccardo's deadline.

The single largest OpenAI infrastructure move of the year, and the first time the lab has publicly owned a silicon design. Per the OpenAI announcement page, the Broadcom investor release, TechCrunch's exclusive, CNBC's on-the-record file, VentureBeat's infrastructure deep-read and the Tom's Hardware tape-out note, the architecture is purpose-built for inference: kernels, memory movement, networking and serving patterns are tuned around the workloads frontier serving actually runs (long-context KV traffic, tool-call control flow, multi-agent fan-out), not a general training-and- inference ASIC. The nine-month-design-to-tape-out cadence is the eye-catching number — Broadcom's release frames it as the fastest reticle-scale high-performance ASIC bring-up on the public record, accelerated by OpenAI's own models being used inside the design loop. Two reads. (1) The multi-generation platform framing is the clearest signal that Jalapeño is the first step of a roadmap, not a hedge — Broadcom says the partnership is now sized to ship gigawatt-scale data centres with Microsoft and other partners beginning in 2026, the same window the Jalapeño first deployment is scheduled for. The inference-tier capacity question that has gated GPT-5.6, the Codex background-agent fleet and every OpenAI consumer surface since Apr now has a first-party answer in the timeline, even if it doesn't deliver wafers tomorrow. (2) The GPT-5.3-Codex-Spark-already-running detail is the under-the-radar tell: OpenAI is not framing Jalapeño as a future-tense bet — engineering samples are running production-class model code at production-target frequency and power in the lab, and the model name picked for the demo is a Codex-tuned Spark variant, not a flagship. The harness layer is the surface getting validated.

The first congressional deadline of the freeze meets the first Polymarket restoration date. Per the Liccardo office's Jun 18 bipartisan letter to Commerce Secretary Howard Lutnick — co-signed by Obernolte, Lieu and Franklin — the asks are specific: (a) the legal authority being used (the letter formally names 14 C.F.R. § 744.22(b)'s "is informed" mechanism under the Export Control Reform Act of 2018's "emerging and foundational technologies" clause), (b) the technical evaluation that drew the line, (c) the review process, and (d) the restoration / licensing criteria — all in writing by Jun 26. Two reads. (1) The letter-on-the-record shape turns Commerce's private-channel directive into a discoverable object: even if Lutnick answers under Touhy protections, the statutory hook argument now has a House-of-Representatives file attached to it, which is the same surface Anthropic's Jun-10 Senate Banking letter covered on Day 13 exited onto. (2) The Polymarket-Jun-26-mid-teens close is the cleanest external read on whether the Tom-Brown-replaces-Dario delegation move that lifted next-week's restoration contract is being priced as this-week: it isn't. The Jul-1 book is still where the weight sits. The flagship-availability negotiation is now publicly-receipted on both sides — industry-side in a Senate Banking hearing letter, congressional-side in a Commerce-direct demand letter — and Day 14 is the day the first deadline ticks past.

The harness team breaks the freeze on the stable channel while the alpha train keeps stacking. Per the openai/codex release pages: rust-v0.142.1 at 00:36 UTC is a tightly scoped Windows-proxy patch from @canvrno-oai via PR #26708; rust-v0.142.2 at 07:32 UTC is the full maintenance roll-up — MCP tool-search is now the default tool-discovery mode (with compat fall-back for older models), the macOS auth client joins the Windows one in respecting system proxy / PAC / WPAD, plugins gain a dark-mode logo hook through local manifests and remote catalogs, and apps can render richer safety-buffering UI using server-provided visibility and faster-model metadata. Two reads. (1) MCP-tool-search-default is the meaningful one for harness consumers: it flips the contract for how Codex walks an MCP server's surface area from list-everything-up-front to search-on-demand, which is the same architectural lift the MCP 2026-07-28 RC made at the protocol layer — Codex is now harness-default on the new shape. (2) The fact that two separate stables ship in seven hours after a twelve-day stable freeze is the clearest signal that the v0.143 alpha train is shielding a separate release branch from user-facing churn — Codex is running a two-track ship cadence now (stable patch + alpha pre-release), not a single-track one.

The alpha half of the new two-track cadence stays at full speed. The v0.143.0-alpha.10 tag landed on Sat Jun 20, alpha.15 on Wed Jun 24 19:41 UTC, and the alpha.16–alpha.25 window now stacks ten more in ~22 hours — a per-hour pre-release rate above anything OpenAI's Rust harness has run on the public record, with v0.143 stable still uncut twelve days after the window opened. Two reads. (1) Holding v0.143 stable off the channel while v0.142.x takes the user-visible patch traffic is a release-engineering choice that only makes sense if the v0.143 ship has a single headline surface being gated — GPT-5.6-class backend routing, a model-tier swap, or a first-party computer-use mode on the harness side. Twenty-five alphas is not churn; it's integration. (2) The cadence itself is the productisation tell: the Mythos-class capability gap that Anthropic opened on Jun 9 with Fable 5's ship — the gap that took Mythos 5 out of public hands three days later — is being closed by harness-side velocity at OpenAI, not by a single GPT-5.6 ship.

The matching Claude Code move, shipped the same day Codex breaks its stable freeze. Per the anthropics/claude-code changelog entry for v2.1.193: autoMode.classifyAllShell is the headline — the setting forces every shell command through the classifier, removing the heuristic-on-suspicious bypass that has historically let deterministically-safe-looking commands skip the auto-mode model call. The claude_code.assistant_response OTel event finally puts a structured, redactable surface under model output in production telemetry, where prior versions left consumers reading raw transcripts to reconstruct what was said. Two reads. (1) classifyAllShell is a pricing move as much as a security move: the auto-mode classifier is a model call, and routing every Bash tool invocation through it changes the per-action cost shape of any agent the harness runs. Putting the option behind an opt-in setting is Anthropic letting enterprise policies eat the budget on their own terms. (2) The OTel log event paired with auto-mode denial reasons in /permissions recent-denials is the cleanest read on the compliance surface Anthropic is now building around the harness — the same surface the Claude Tag Slack rollout on Day 13 requires under enterprise admin scoping.

The cleanest read this week on where the computer-use surface consolidates. Per the blog.google announcement, the TechTimes file and the Cybersecurity News write-up, a single Gemini 3.5 Flash agent can now see a screen, look something up on Search, and interact with a map without routing requests between models — the same model that already exposes function-calling, Search grounding and Maps. Two reads. (1) The built-in-not-bolt-on shape is the meaningful product move: the Oct-2025 Gemini 2.5 computer-use model shipped as a standalone surface, which forced agent authors to multi-model around it; making computer use a default tool inside the same Flash model developers already call collapses the routing layer and changes the cost shape of a browser-driving agent at the API tier. (2) The 78.4 vs 78.7 OSWorld gap at $1.50/$9 vs $5/$30 is the under-the-radar tell: Google is comfortable shipping an approximately-parity computer-use benchmark at roughly a fifth of the GPT-5.5 input price and a third of the output price, which is the commodity-tier pricing move the OpenAI harness train and the Anthropic Slack/Tag surface have been ducking on the computer-use tier.

The single largest vertical-AI-agent round of the week. Per the PRNewswire release, the FierceHealthcare file, HITConsultant's investor read and citybiz's valuation confirmation, the $1.2B mark is a unicorn-conversion step for a company whose 190M-patient-voice- interactions footprint is the cleanest production-scale reference for an enterprise voice-agent platform in U.S. healthcare. The "agentic operating system for the patient journey" framing signals the next product surface: expansion from patient access (scheduling/intake) into document processing, medication refills and payments — the same Box-Skills-class action-on-content rim covered Day 13, applied to a regulated vertical with 20× ARR growth as the financing thesis.

Per the Taktile release, the Axios Pro Rata file, the FinTech.Global write-up and the Yahoo Finance deal-page, Goldman Sachs Alternatives leading the round is the meaningful detail: the $90M-projected-savings-at-one- insurer data point is the kind of single-customer ROI receipt growth-equity tickets ride on, and the round positions Taktile as the decision-layer incumbent for U.S./EMEA/LATAM banks and insurers ahead of the EU-AI-Act-Aug-2-cutover the spec community has been counting down to. The $110M-out-of-$184M-total shape is the read on how growth-tier investors are sizing the agentic-finance-decisions opportunity now that banks-and-insurers are the actual deploying customer, not an NDA-bound reference.

The governance-layer bet of the week. Per Fortune's exclusive, the Techmeme roundup, the Finsmes file, the SaaSNews deal record and the Crypto Briefing write-up, Felicis leading with Khosla Ventures already in positions the round at the intersection of enterprise-AI-procurement and agent-governance-and-observability — the same surface OpenAI-MCP-OAuth-hardening and Anthropic's per-channel-data-and-tool-scope in Claude Tag are now both shipping on the model-and-harness side. The "every-available-dollar" quote from Khosla is the investor-side equivalent of the Claude-Tag-65%-of-product-team-code disclosure: a credibility deposit at the governance moment of the vertical, not the experimentation moment.

The evals-and-simulation bet that closes the week's vertical-agent funding stack. Per the PRNewswire release, SiliconANGLE's production-pipeline read, TheNextWeb's product file and the Finsmes deal record, Coval's pitch is the production-side counter to Assort Health's patient-journey deployment: simulate the agent against thousands of synthetic customer-call branches before it ships, observe and label what happens in production, and review the regressions between releases. Zoom and Deepgram as named customers map cleanly to the voice-AI surface that OpenAI's GPT-Realtime-2 / Translate / Whisper GA earlier this month opened up at price/$M-token-tier. Norwest leading positions Coval as the evaluation-platform incumbent for the realtime-voice surface the way Snorkel's $100M Series D positioned it on the text-agent rim.