OpenAI ships ChatGPT Work, Codex becomes ChatGPT, GPT-5.6 goes public — and Cerebras carries them at 750/1000 tokens/sec
— forty-eight hours after Anthropic put Cowork on phones, OpenAI answers with a super-app triple, GPT-Live turns ChatGPT full-duplex, Cognition's SWE-1.7 lands on the same wafer-scale silicon that hosts GPT-5.6 Sol, and Illinois signs the first US frontier-model safety law with 72-hour incident reporting.

The first head-to-head answer to Anthropic's Cowork plural-surface move — and the outcome-oriented + hosted sites combination is the story. Per the 9to5Mac, MacRumors, GVWire, TestingCatalog, iPhone in Canada, Reuters and Digital Applied reads of the launch: ChatGPT Work ships as the agent tier whose job description is "finish the work," and the same launch day sees OpenAI collapse Codex, ChatGPT and hosted sites into one surface. Two reads. (1) A ChatGPT Work agent that connects natively into Slack, Teams, Drive and Salesforce is the exact work-execution surface Anthropic described on Jul 7 when Cowork shipped on mobile with the disclosure that ∼91% of Cowork usage is not coding — the two vendors are now shipping the same shape (agent + phone + desktop + connectors) on consecutive weeks, which means the "general-purpose work-execution" framing is now the consensus product thesis rather than a land-and-widen from either camp. (2) Hosted sites inside ChatGPT Work is the move Cowork doesn't yet make and the buyer signal is self-serve publishing — the same buyer that would have paid Vercel or Netlify for a marketing page now gets it as an agent side-effect, which puts the frontier lab inside a category (web hosting) where no frontier lab has shipped a product before, and positions Plus and Business availability (rolling in days) as the consumer distribution curve for a hosting primitive.

The brand collapses to match the product collapse — and the direction of the merge is the story. Per the MacRumors, 9to5Mac, TestingCatalog and Digital Applied reads: Codex shipped as an autonomous coding surface on the desktop; the Jul 9 rename means the coding-agent client is now the general-purpose client, not the other way around. Two reads. (1) The direction — Codex becomes ChatGPT, not ChatGPT gains a Codex mode — matches the Anthropic disclosure on Jul 7 that Cowork is ∼91% non-coding: at both labs the agent harness that started as a coding tool is now the work-execution surface, and the labels are catching up to the usage. (2) A single ChatGPT desktop client that houses both the Codex agentic-coding surface and the ChatGPT Work autonomous-work surface means the ChatGPT brand now owns the desktop distribution slot the same way Claude Desktop owns it on the Anthropic side — which is where Cowork and ChatGPT Work both live as clients, not as websites, and where the next MCP-native integrations will land first.

Materially new development on the CAISI clearance (yesterday's item 06) — the public GA is the proof-of-process, and the Cerebras speed number is the second story on the same day. Per the Bloomberg, Fortune, CNBC transcript, Washington Post, Engadget, Neowin, Simon Willison and ValueAdd Pulse reads: Altman's on-camera framing — "many changes", "collaborative back and forth" — is the first frontier CEO admission that the voluntary gate materially shaped the released model rather than simply blessing what was shipped, and the 750 tok/sec Cerebras number turns the specialty silicon layer into the latency story the same week the ChatGPT Work agent surface needs sub-second round-trips to feel like a supervisor rather than an inbox. Two reads. (1) The $1/$6 Luna floor and the $5/$30 Sol ceiling put the GPT-5.6 family below Fable 5's $10/$50 credit meter on every axis, and Fable 5 is on a Jul 12 billing cliff (yesterday's item 07) — the same Anthropic Max cohort weighing the credit meter now has an Opus-class-ish option at a fifth the headline price, delivered at 10× the tokens-per-second. (2) The "many changes" framing lands after the Anthropic Lutnick letter (Jun 30) established the same pattern for Fable 5 — voluntary submission, Commerce review, WH refusal to call the outcome a "permit" — and the Illinois SB 315 signature four days earlier (item 12) means the state layer is already ready to codify what the voluntary framework just proved.

The first full-duplex voice shape from a frontier lab — and the delegation to GPT-5.5 is the story. Per the TechCrunch, MLQ, MacRumors, HotHardware, Dataconomy and TestingCatalog reads: the duplex architecture removes the ASR → LLM → TTS stitch that made Advanced Voice feel turn-based, and the ability to continue the conversation while GPT-5.5 handles a background query is the first time a voice agent gets a frontier-model tool call without the conversation stalling. Two reads. (1) A duplex voice surface on the same ChatGPT desktop client that now houses Codex and ChatGPT Work is the third agent modality to converge on the same distribution — the coding-agent, work-agent, voice-agent triple is now one client, not three products, and Anthropic Advanced Voice plus Cowork Desktop is the mirror answer that Anthropic hasn't yet published. (2) A Free-tier default of GPT-Live-1 mini lands the full-duplex surface on the same tier that gets Copilot app (yesterday's item 02) — the pattern is becoming visible: frontier models (GPT-5.6 Sol, Fable 5) stay credit-metered, but the experience layer (voice, agent, harness) lands at Free to keep the funnel attached.

The first European frontier lab to publish a robotics model in the agent-era positioning — and the compact / simulation-only / single-camera combination is the story. Per the Bloomberg, Mistral release, The Decoder, Silicon Republic, TestingCatalog and Investing.com reads: Robostral Navigate is a vision-language-action model on ∼8B parameters that turns "go to the loading dock" into a chain of low-level motor commands from a single RGB camera, and the fact that it beats depth + multi-cam systems on R2R-CE is the first data point that frontier compression translates into hardware BOM reductions in embodied deployments. Two reads. (1) A hardware-agnostic model deployed onto Airbus and BMW pilots skips the humanoid narrative that Figure, 1X and Optimus anchor — the frontier bet is navigation as a substrate for industrial AGVs rather than general humanoid autonomy, which is the same bet the Bezos $41B Physical AI commitment pointed at earlier in the year. (2) A Mistral push into robotics the same week OpenAI converges the desktop and voice surfaces and Anthropic converges the desktop and mobile surfaces means the three frontier labs are now expanding on orthogonal modality axes rather than fighting for the same modality slot — a polycentric shape the buyer will notice first inside logistics and manufacturing accounts.

The first Devin-house frontier-tier coding model that lands on the same Cerebras silicon OpenAI is using for GPT-5.6 Sol — and the RL on top of RL training recipe is the story. Per the Cognition blog, Winbuzzer, vktr, TechTimes, ExplainX, ChatForest and Mervin Praison reads: SWE-1.7 starts from Kimi K2.7 Code (Moonshot's already-RL-post-trained 1T-total / 32B-active MoE) and adds Cognition's own globally-distributed RL pipeline across four data centres on three continents, combining Cognition's GPUs with Fireworks inference. Two reads. (1) The 81.5% Terminal-Bench 2.1 and 77.8% SWE-Bench Multilingual numbers are the first third-party read on how far open-weight bases can be pushed by proprietary RL fine-tuning: K2.7 Moonshot pricing is $0.95 / $4 per M token, and shipping a model with near-frontier Terminal-Bench performance on top of it means the base + RL stack is now the cheapest path to a frontier-adjacent coding tier, which is what Cursor Composer 2.5 already proved on a different base and what Grok 4.5 just proved on Cursor data. (2) The 1000 tok/sec Cerebras throughput pairs with GPT-5.6 Sol's 750 tok/sec on the same silicon (item 03) — Cerebras now carries two frontier-tier coding deployments simultaneously, and the "is this a >10× latency win or a marketing number?" question gets answered by developers at both Devin Desktop and ChatGPT Codex inside the same 72-hour window.

The Codex stable cut that ends the Mythos-restoration freeze — and the Bedrock GPT-5.6 first-class-model row is the story. Per the Releasebot, Codex changelog, Havoptic and Releases.sh reads of the tag: the v0.143 pre-release train stacked past thirty-one alphas (previously covered) and the stable cut lands the same day GPT-5.6 clears CAISI and SWE-1.7 lands on Cerebras. Two reads. (1) A first-class Bedrock GPT-5.6 model row inside Codex means the AWS-hosted version of the flagship is now the developer default at exactly the moment ChatGPT's hosted version is also going public — the Bedrock ↔ ChatGPT distribution symmetry OpenAI shipped last quarter is the shape a choose-your-cloud buyer wanted, and the same shape Anthropic offers via Bedrock, Vertex, Foundry and native Platform today. (2) macOS + Windows system-proxy support is the enterprise-desktop stamp that Codex had been missing — the same PAC / WPAD traversal anthropics/claude-code shipped earlier in the cycle, which closes the last corp-network gap between the two harnesses inside enterprise laptops.

The specialty inference layer becomes a first-order frontier vector — and the "OpenAI workloads explicitly named" phrasing is the story. Per the Cerebras IR release, Euronews, Yahoo Finance, Reuters, Seeking Alpha and Invezz reads: this is the same WSE-3 substrate hosting GPT-5.6 Sol and SWE-1.7, and the 200 MW / end-2027 target is roughly what a single hyperscaler announces at a country scale — Cerebras is doing it at a bloc scale on frontier-model workloads. Two reads. (1) A France + Nordics anchor gives the European buyer a sovereign-friendly latency floor for OpenAI workloads the same week the EU Action Plan on Cybersecurity + AI (yesterday's item 11) tapped ENISA to build "secure access" to advanced AI systems — the infrastructure answer to the procurement framework is landing before the framework has teeth. (2) The Nordic selection lines up with the hydro and district-heat grid attributes that Meta, Microsoft and Google have chased for years — the wafer-scale cooling profile favours cold-climate sites, which turns the Nordic power/cooling story into a capex multiplier that H100 deployments do not get to use at the same scale.

The own-your-agents buyer thesis gets its own unicorn — and the Radical + Nvidia + Intel + Dell cap-table combination is the story. Per the TechCrunch, PYMNTS, Prime Intellect blog, Yahoo Finance and aVenture reads of the round: Prime Intellect sells the "stack that Anthropic and OpenAI don't sell" — RL frameworks, eval harnesses, distributed training compute — to enterprises that want to train the agent themselves. Two reads. (1) A $100M ARR run rate on a company selling "build your own Claude Sonnet-shaped thing" means the own-stack buyer thesis is now revenue-real rather than fund-real — the Ramp / Zapier / Flapping Airplanes customer list is heavy on internal-workflow agent operators rather than customer-facing agent operators, which is the same segment Norm Ai (yesterday's item 08) and Microsoft Frontier (yesterday's item 09) are also chasing inside asset managers, banks and enterprise IT. (2) A cap table with Nvidia, Intel and Dell together signals the hardware layer sees own-stack enterprise training as a compute-buying thesis: if Ramp-shaped buyers keep training their own agentic systems on Prime Intellect, the silicon pull-through is directly enterprise-hosted rather than routed through a frontier lab — which is why three hardware cap-table participants showed up simultaneously.

The first frontier-agent startup that runs its own fundraise through its own agent — and the "130 investors, no flights" mechanic is the story. Per the Bloomberg, TechCrunch and The Next Web reads: SivaClaw takes an investor question shape (data-room requests, technical DDs, market-sizing memos) that a partner-track venture associate would normally do, and Lyzr doubles the Mar 2026 valuation in four months on the strength of the demo. Two reads. (1) An agent-run fundraise closing at $500M is the first marketing artefact that turns "we build agents for the enterprise" into "we use our own agent to acquire capital" — the same shape OpenAI uses when it publishes ChatGPT-drafted communications, and the shape Norm Ai (yesterday's item 08) uses when its Norm Law AI-native firm bills clients. (2) The Accenture-led Mar 2026 round's $250M valuation doubling in four months is a data point on how fast enterprise agent valuations are repricing on execution: the Prime Intellect (item 09) and Norm Ai (yesterday) and Lyzr unicorn set all crossed the $1B-adjacent line inside a single trading week.

The offensive-security tail from Five Eyes Jun 22 ("months not years") gets its first documented artefact — and the 31-second recovery timing is the story. Per Sysdig, TechCrunch, BleepingComputer, Dark Reading, Business Standard, Android Headlines and Security Affairs reads: JADEPUFFER is the first public incident where an agent ran the full MITRE ATT&CK chain without human supervision of the attack execution (a human still provisioned the infrastructure and picked the target). Two reads. (1) A Langflow CVE as the initial-access vector is the first time a public ransomware operation started at the agent-orchestration layer itself — the same layer that enterprises are deploying at scale inside Anthropic Cowork, ChatGPT Work (item 01) and Devin Desktop (item 06) — which widens the attack surface to include the agent platform as infrastructure, not just the web application. (2) The ENISA "machine-speed threats" framing published on Jul 7 and the EU Action Plan on Cybersecurity + AI (yesterday's item 11) both point at exactly this shape — a 31-second agent-recovery loop is the empirical observation that machine-speed incident response becomes a vendor requirement, and Illinois SB 315's 72-hour incident-reporting clock (item 12) starts on the same tape.

The first state-level answer to the same Jun 2 EO voluntary framework CAISI just used to gate GPT-5.6 (item 03) — and the 72-hour clock is the story. Per the Illinois Governor release, Capitol News Illinois, Washington Post, Inside Global Tech and Governing reads of the signing: SB 315 is the first US law to statutorily require what the EO framework asks voluntarily, and the 72-hour reporting window is tighter than the SEC cyber incident-disclosure rule (4 business days). Two reads. (1) A California + New York + Illinois trio at ~40% of the US AI market gives the state layer de facto national coverage — the same Brussels-effect shape the EU AI Act's Aug 2 fines-live deadline created for Europe now materialises inside the US, and any frontier lab that ships Fable 5, GPT-5.6, Gemini 3.5 Pro or Grok 4.5 to US buyers has to file the same catastrophic-risk paperwork three times in three states before the federal framework has finished consulting. (2) A 72-hour incident-reporting clock lands on the same Jul 7 tape Sysdig publishes JADEPUFFER (item 11) — a real AI-agent ransomware operation where the attack recovered in 31 seconds. The clock is now calibrated against observed threats rather than hypothesised ones, and the frontier labs whose API keys were harvested by the agent (per Sysdig) are on the reporting side of the same law.