The AI Wire ●

During the DevDay keynote, OpenAI reset every Codex user's usage allowance worldwide — at the 52:15 mark of the keynote, per a timestamped transcript. Each attendee also left with a limited-edition ModRetro Chromatic handheld built with Codex.

The reset came the same day Codex moved into the cloud with reusable team environments, a voice-steered CLI with an /agents view, Code Review in the desktop app, and Codex Security Cloud scanning repositories on a schedule. The CLI's default model is now GPT-6.1 Sol at medium effort.

This is the practical side of DevDay's pricing shake-up: Pro 200 reopened with new usage math, a $500 Pro 500 tier with Ultrafast access, and the Ultrafast speed tier itself (up to 8x faster in Codex, 6x in the API). Whether the reset becomes a per-keynote ritual — or a one-off sugar rush — is worth watching next quarter.

OpenAI announced a batch of Codex updates at DevDay aimed at making the coding agent useful beyond a single laptop. Reusable cloud development environments let developers run Codex from a computer, a phone, or the cloud, starting tasks faster on shared, persistent workspaces with approved settings and permissions instead of treating each cloud task as an isolated sandbox.

The Codex CLI now supports starting and directing tasks by voice, and a new /agents view makes it easier to delegate and track multiple tasks at once. Editing prompts, resuming sessions, worktrees, and a cleaner terminal UI for long sessions round out the developer-tooling refresh.

Codex also gets a new code-review experience inside the ChatGPT desktop app: developers can read summaries, explore changes, and ask Codex about issues before commenting on GitHub pull requests or GitLab merge requests. An automatic mode lets Codex do a first review pass while the developer steps away — part of OpenAI’s push to embed agents across the software workflow.

At DevDay, OpenAI moved deep into workplace software territory with ChatGPT Space, a shared workspace where coworkers, ChatGPT, and their Dots agents collaborate on live projects. “Your pages and files all live together in Space, like they would in a drive,” Altman said during the keynote. “But spaces feel alive.” Space is available on Pro, Business and Enterprise plans on web and desktop.

Alongside it came Pages, a new document format “built for human and agent collaboration” that can write, research, generate charts, create images, and visualize information — launching in Q4 2026. OpenAI also showed collaborative slides that can be generated by talking in ChatGPT and co-edited by people and agents, exportable to PowerPoint and Google Slides.

The launch puts OpenAI in direct competition with Microsoft's Office suite even as the companies partner on Azure and Agent 365 integration. Also in the productivity push: a Meetings plugin for calendar-connected notes, @ChatGPT integrations for Slack and Teams, and a plugin-sidebar system with 16 launch partners sharing plan allowances.

OpenAI gave developers a way to live permanently inside ChatGPT at DevDay on September 29: plugin extensions give a plugin a dedicated home in ChatGPT's sidebar rather than a one-off tool call. Developers can also build interactive panels users work in alongside the conversation, plus viewers for whatever file types their product uses — available on every ChatGPT plan.

A new Plugin Creator tool handles the build side, submission now runs through a redesigned flow with clearer feedback, and improved ranking surfaces plugins in the directory and live conversations. Plugins can trigger automations off events in a connected app — for example, watching a project board and drafting a plan while the user is away — via OpenAI's new support for the proposed MCP Events specification.

Business, Enterprise, Healthcare and Edu customers can host supported plugins inside the ChatGPT Sites they build, sharing one connected app across a workspace with per-member permissions. TechCrunch read the move as a direct challenge to app-store distribution, now that ChatGPT counts 1.2 billion weekly users.

OpenAI and Amazon launched Bedrock Managed Agents at DevDay — powered by OpenAI, the product lets companies build OpenAI agents that run entirely inside Amazon Web Services, with data never leaving AWS. The product is in limited preview, per the AWS product page.

It wraps the newly public-beta Agents API — hosted execution, computer use, memory, multi-agent control, tool search and context compaction — so AWS-native teams can use OpenAI's agent stack without routing data outside their cloud. For teams whose data and procurement already live in AWS, that removes the usual reason not to try OpenAI agents.

Altman also claimed 'more than 99 percent' platform uptime and a 45% cut in time to first token. The AWS partnership lands amid a broader corporate shift toward cheaper open-weight models (FT, Sept 28) — OpenAI is keeping its agent stack sticky inside the clouds enterprises already pay for.

Bloomberg reported Tuesday that OpenAI is in early talks with investors to raise at least $30 billion in a new funding round at a valuation of roughly $1.4 trillion, not including the money raised. The discussions are investor-led and still early, with terms that could change, and OpenAI did not immediately respond to comment requests.

On October 6 Bloomberg followed up with named investors: several UAE funds including MGX in Abu Dhabi are discussing forming a consortium that could contribute up to $10 billion total as core investors, and BlackRock is in discussions to participate alongside it. Existing investors including the University of California endowment, Thrive Capital, and Andreessen Horowitz have also been engaged. OpenAI is offering the ~$1.4 trillion pre-money valuation as a fixed price with no determined lead investor, and has postponed IPO plans until at least next year, citing a focus on AI safety efforts.

The raise would serve as bridge financing ahead of an anticipated IPO and follows the $122 billion round in March at an $852 billion valuation, which was supposed to be the last private raise before going public. Still in active talks with final terms and investors not settled — treat as rumor until OpenAI or the investors confirm.

On October 9 Bloomberg confirmed the fundraising is still live: OpenAI told investors it expects to reach or exceed $70 billion in annualized revenue by the end of 2026, driven largely by enterprise growth, and shared the figures as part of the $30 billion-plus raise discussions at the ~$1.4 trillion pre-money valuation. September annualized revenue was roughly $50 billion. OpenAI pushed its IPO plans to at least next year; Anthropic, whose annualized revenue hit $65 billion by the end of July, is expected to IPO as soon as November.

OpenAI, Google and Anthropic have agreed to testify before the New York City Council at an AI hearing on October 5 — but only after council leadership threatened to use its subpoena power. Meta was the only company to agree before the threat was issued; SpaceXAI never responded and was formally ordered to attend.

Council Speaker Julie Menin said the companies employ tens of thousands of New Yorkers and "owe it to their workers and, above all, to the public to come before the council, answer our questions, and provide input on our proposed legislation under oath." The hearing will examine the dangers of rapidly advancing AI after reports citing tens of thousands of safety incidents.

The episode fits a pattern of escalating government scrutiny: the FTC has opened an industry-wide probe into rogue agents, and OpenAI and Anthropic declined to appear at an Australian Senate AI hearing this week, citing the short notice.

Meta launched Muse for Small Business on Tuesday, an agentic AI assistant aimed at small business owners — from creating ad campaigns and analyzing financial performance to sorting and answering incoming mail. Meta is pitching it as an AI co-worker for owners who'd rather be on the shop floor than in the office, with early testers like a grocery owner and a clothing-brand founder quoted in the announcement.

The agentic model can work in the background while users do other things and take action on their behalf — but Meta stressed it will not take action without the user's permission. Business owners are prompted to name their Muse on sign-up, and 35 business owners were invited to Meta's Los Angeles campus in early September to test and refine the product.

This extends the Meta One push — AI features across Instagram, Facebook and WhatsApp — into a dedicated SMB product. With OpenAI launching its always-on 'dots' agents the same day, the two are racing head-on for the personal-agent slot on consumers' phones and PCs.

Cognition has cut Devin's prices across the board: 30-40% cheaper in Fusion and Normal modes, 15-20% cheaper in Ultra mode, and up to 70% cheaper for Devin Review. The company credits smarter model routing and a more efficient harness — batching commands, running independent actions in parallel, heavier prompt caching — rather than any reduction in quality.

The price cut lands alongside a benchmark claim: Devin Fusion, the lead-and-sidekick harness that pairs a frontier model for planning with a cheaper model for execution, scored 68.8 on the Extended portion of FrontierCode 1.1 at an average of $0.60 per task, which Cognition says ranks first. Fusion draws on Cognition's own SWE-2 model alongside Opus 5.5 and the GPT-6 family, and the company says it is up to 39% more efficient than other harnesses across major coding benchmarks.

The timing is notable: the cuts come days after Bloomberg reported Cognition crossed $1 billion in annualized revenue — roughly double its run rate from four months earlier — with customers including GE Aerospace and Rivian. Devin bills on usage-based Agent Compute Units rather than a flat subscription, so heavy users will feel a move like this directly in their monthly spend.

Anthropic confirmed a partial outage across its Claude platform on Tuesday, September 29, after a disruption began shortly before 10:00 a.m. ET. Elevated error rates affected Claude.ai — including the desktop and mobile apps — along with Claude Code, Claude Cowork, and eventually the Claude API, with users seeing failed requests, error-loading conversations, and forced sign-in loops. Downdetector recorded over 18,000 problem reports at peak, with reports concentrated in the US and India.

Anthropic opened an incident on its status page at 14:21 UTC, applied a partial mitigation at 14:41 UTC, then hit a second failure that took out single sign-on, Sign in with Apple, new chats, voice conversations, purchases and file uploads. By 14:59 UTC (10:59 a.m. ET) the company said all services were operating normally — roughly an hour after the 14:00 UTC start — and the incident moved to monitoring at 15:11 UTC. Anthropic warned that some messages sent between 14:00 and 14:59 UTC may not have been saved, and never disclosed a root cause. Signed-in users were told not to log out, to avoid authentication loops.

It was the 13th incident logged on Anthropic's status page this month, with September uptime at 99.70%, per Unite.ai. Spotify suffered its own outage at the same time — Downdetector reports peaking above 19,000 — though there is no sign the two were linked. 9to5Google noted limited, inaccurate reports of ChatGPT and Gemini being down too; neither OpenAI nor Google confirmed issues.

The timing is awkward: the outage landed one day after Anthropic released Claude Sonnet 5.5, and on the same morning as OpenAI's DevDay keynote — a day when the frontier labs were hoping to look their most reliable.

Anthropic confirmed a third model in the Claude 5.5 family: Claude Haiku 5.5, aimed at high-volume and cost-sensitive use, due in the coming weeks. It rounds out the family — Opus 5.5 for the hardest work, Sonnet 5.5 for well-scoped everyday tasks, Haiku 5.5 for scale.

The family strategy is clear: bring flagship-level capability down into daily-work economics. Sonnet 5.5 already scores near Opus 5.5 on long-horizon knowledge-work benchmarks at Sonnet 5's list price, with up to 30% less per task; Haiku 5.5 should push the cost floor further still.

Watch the pricing when it lands — Haiku models have historically been where Anthropic fights for agentic-loop volume, and Sonnet 5.5's independent cost findings (above) make the real per-task economics worth checking before switching.

OpenAI product lead Tibo Sottiaux announced the reopening of new $200/month ChatGPT Pro sign-ups, timed to DevDay, after a three-week pause that began September 10. The $200 plan was OpenAI's highest usage tier — 20x the usage of Plus at launch — and was paused because Pro users were the highest per-user strain on capacity. Note: after DevDay, NEW Pro subscribers get 10x the Plus allowance; existing subscribers keep 20x.

The reopening comes with a new usage calculation and a key commitment: OpenAI says it will not reintroduce the five-hour usage limit removed in July, so subscribers can spend their full weekly allowance whenever they want. Tibo also promised that subscribers would get 'more work than a month ago' as model efficiency gains are passed down through API price reductions.

Questions from subscribers remain unanswered: whether 30-day notice was required for plan math changes, grandfathering of existing terms, and whether the reopening amounts to half the previous API-dollar allowance. Watch for the details in the DevDay keynote and OpenAI's updated help-center docs.

Anthropic's Frontier Red Team published its assessment of Zhipu AI's open-weight GLM-5.3 on September 29, and the numbers are striking. On ExploitBench — exploiting known vulnerabilities in Chrome's V8 engine — GLM-5.3 built end-to-end exploits in 50 of 410 attempts, against 56 for Anthropic's own Claude Mythos Preview. On Anthropic's binary-exploitation benchmark, GLM-5.3 achieved a full control-flow hijack in 4% of trials versus 6% for Mythos. Earlier models — Claude Opus 4.6, GLM-5.2, Kimi K3, DeepSeek-V4.1-Flash — scored zero.

The report's sharper point is about safeguards. GLM-5.3 refused direct malicious requests in Anthropic's simulation, but compliance rose to 64% with a deceptive cover story, 92% with prefilled reasoning, and 100% after the open weights were modified to strip refusals. In one session the model found several previously unknown browser flaws and chained them into a webpage that escapes the sandbox and reads files from a visitor's computer. NIST's Center for AI Standards and Innovation called GLM-5.3 the most cyber-capable open-weight model to date on September 17.

This is the genuine-significance exception to the Chinese-lab rule: a safety finding published by a major US lab, with industry-wide implications for open-weight releases. Anthropic's framing is pointed — advanced exploit capability has now spread beyond the vetted-defender ring it built around Mythos through Project Glasswing.

Anthropic expects to spend at least $518 billion over a decade building AI infrastructure with six partners, according to its confidential IPO prospectus reviewed by Reuters — a plan that ranks among the largest AI buildout commitments on record. About 80% of the sum is non-cancelable or payable regardless of usage.

The filing commits at least $111.1 billion to Google and $110 billion to Amazon under long-term service obligations over the next seven to ten years "regardless of usage," plus $31.4 billion to Microsoft and roughly $161.2 billion in largely non-cancelable Broadcom equipment leases. Only the xAI agreement, worth up to $84.5 billion, can be canceled on 90 days' notice.

The prospectus also reveals the numbers behind the planned November IPO: nearly $4.6 billion in 2025 revenue (a twelvefold jump), a $42 billion net loss including a ~$34 billion non-cash financing charge, $20.28 billion in cash and short-term investments, and a possible $2 trillion valuation. Anthropic says access to computing power is becoming the key constraint on AI development, with future demand "limited principally by the availability of compute."

GitHub added Anthropic's Claude Sonnet 5.5 to Copilot the same day the model launched, September 28. It is generally available to Pro, Pro+, Max, Business, and Enterprise users across VS Code, Visual Studio, the Copilot CLI, JetBrains, Xcode, and github.com, billed at provider list pricing under usage-based billing.

For developers, that puts the headline efficiency story directly into their editor: Anthropic says Sonnet 5.5 generates output more than 30% faster than Sonnet 5 and can cut cost per task by up to 30%, while per-token prices hold at $2 per million input and $10 per million output tokens. It scored 70.6% on Terminal-Bench 4.0, ahead of Opus 5.5's 66.4% at its highest effort setting.

It is also the first Sonnet release to ship with the cyber safeguards and fallback protections Anthropic previously reserved for its most capable models, plus new safety classifiers the company says block reasoning-extraction attempts used in distillation attacks.