https://x.com/i/article/2109197366976778240
OpenAI and Anthropic this week: GPT-6 for all, Haiku 5.5, UV sky map, Cyber Mission (Week 41, 2026)
OpenAI started shipping one Codex and Work improvement a day for 28 days, or a full usage reset on days without one, so far bringing GPT-6 Astra and GPT-6.1 Sol about 50% faster by default, free Auto-review, a rate limit reset for everyone, and a banked reset for paid accounts at 40 million Codex and Work users, rolled out GPT-6 with Intelligent UI to everyone in ChatGPT with GPT-6 Sol on paid plans and GPT-6 Luna on Free and Go, Ultrafast mode for GPT-6.1 Sol in the API, Codex, and ChatGPT Work at up to 8x the speed of Sol Standard with steering of running Codex tasks now instant so you can course-correct in real time, the Decisions API in public beta for all developers at $0.10 per million input tokens, up to 10x faster than GPT-6 Luna through the Responses API, three API usage tiers named Build, Launch, and Grow with the top one unlocked at $500, text watermarking for the EU AI Act with the textGrain method set to go open source, a new Codex sandbox on Windows built on Microsoft Execution Containers, composer predictions in Codex in beta for Pro users, dots created from the mobile apps and delegated to Codex, custom MCP servers without developer mode, editable models and reasoning effort for scheduled tasks, Sign in with ChatGPT plan usage for Go subscribers, and added audio uploads on paid plans and workspaces to transcribe, summarize, and ask questions about recordings, will test a new visual ad format in ChatGPT during image generation for Free and Go users starting later this month in the US with new conversion, incrementality, and brand suitability measurement partners, previewed College Planner for ChatGPT for Teens, which I'd earlier seen in ChatGPT as College Plan, published new mathematical results from an internal frontier model with Lean formalizations, alignment research on metagaming latents and the LASER sampling method, and a report on two banned covert influence operations, a Russian "false front" one it calls Dark Clark and what appears to be a for-hire Iranian campaign it calls Bogus Bylines, expanded its Atlassian partnership with GPT-6 Astra and GPT-5.6 series models powering agents across Atlassian's platform and Rovo, started a computer use research collaboration with Ironclad where GPT-6 Astra scored 55.0% against 41.6% for GPT-5.6 Sol on 11 contracting tasks, and its research leaders answered a public letter from three safety researchers the company let go last week, Tomek Korbak, Jasmine Wang, and Mikita Balesni, saying an investigation found they violated clear policies on handling sensitive information, that the decision was not about raising safety concerns, and that contracts with third-party safety assessors will be announced in the coming weeks
Anthropic released Claude Haiku 5.5 as its cheapest, fastest, and most capable small model at $0.10 / $0.50 per million tokens for prompts up to 100K with a 1M context window and the first adjustable effort setting on a Haiku model, halved Sonnet 5.5 cache read pricing to $0.10 per million tokens in the API, added monthly Claude Platform API credits of $100 on Max 5x, $200 on Max 20x, and up to $500 pooled on Team, built computer use and browser use toolsets into the Python and TypeScript SDKs, put Claude Managed Agents dynamic workflows in public beta with a lead agent planning work across up to 1,000 agents per run, let in all waitlisted Pro and Max users for Claude Code Projects, took Claude Docs, Slides, and Design out of beta on every plan including Free with the standalone Design site folding into Claude on December 14, put Claude Dashboards and Claude Motion into beta, brought Claude into Google Docs, Sheets, and Slides as a sidebar add-on in beta on paid plans, expanded the Claude Startups program and then paused its Team and API credit offers after 21x as many applications in 48 hours as in the previous five months, with claimed offers staying and unclaimed ones re-reviewed, expanded its Cyber Verification Program into Defense, Red Team, and Specialized access tiers with Claude Mythos 5.1, Opus 5.5, and Sonnet 5.5 under reduced cyber safeguards, launched the Anthropic Cyber Mission with a Critical Infrastructure Defense Program and a free OSS Scanner for open-source projects, committed $150 million over three years to the Genesis Mission for federal science research, published how astrophysicist Brice Ménard built the first complete UV sky map with Claude Science, a 2026 Usage Policy update taking effect November 12 with clearer weapons and surveillance rules and a ban on sustained needless cruelty toward its models, and the first of its new model behavior reports covering four types of unintended actions seen in evaluations and internal use, with live internet access now off for internal evaluations until new monitoring is confirmed, and more
OpenAI is shipping one Codex and Work improvement a day for 28 days or a full usage reset on any day without one, per Tibo Sottiaux, who now leads ChatGPT and Codex, so far bringing GPT-6 Astra and GPT-6.1 Sol about 50% faster by default for subscribers, free Auto-review for everyone signed in with ChatGPT that does not count against plan usage, a rate limit reset for everyone, and a banked reset for paid accounts to celebrate 40 million Codex and Work users
OpenAI will test a new visual ad format in ChatGPT during image generation for Free and Go users, starting later this month in the US with an initial group of advertisers, with ads clearly labeled and kept separate from the generated image, plus conversion data integrations with Hightouch, Tealium, and LiveRamp, geo-based incrementality experiments with Haus, Measured, and WorkMagic, Negative Phrases for qualifying advertisers, and brand suitability pilots with DoubleVerify and IAS
Scheduled tasks now let you edit the model or reasoning effort directly from the Edit task modal
OpenAI is rolling out text watermarking for the EU AI Act opt-in for API customers globally on select models from October 5 and added to eligible ChatGPT and Codex text in the EU over the coming weeks, with the detector limited to approved researchers given weak detection on short, edited, or translated text, and the textGrain method, which matched or beat SynthID for text in OpenAI's tests without meaningful quality loss, set to go open source
ChatGPT no longer requires developer mode to connect custom MCP servers so all users can add them under Plugins, Add, Create custom MCP server
Anthropic expanded its Cyber Verification Program into three access tiers Defense Access for defensive work, Red Team Access for authorized penetration testing and red-teaming by organizations only, and Specialized Access for safety-critical systems reviewed with the US government, each with Claude Mythos 5.1, Opus 5.5, and Sonnet 5.5 under reduced cyber safeguards, folding in Project Glasswing, whose partners found at least 129,000 verified vulnerabilities between April and July 2026, with data retention required for misuse monitoring, and Enterprise Frontier Safeguards arriving later this fall to let eligible organizations keep that data in cloud infrastructure they control
OpenAI published new mathematical results from an internal frontier model in a GitHub repository with Lean formalizations of many proofs, 10 summaries of the model's reasoning, and compute estimates averaging about three hours of ChatGPT Pro thinking per result, following advice from the Institute for Advanced Study's advisory group on mathematics and AI, with funded workshops and a responsible release of the model still to come
OpenAI updated dots in ChatGPT with creation from the mobile app on iOS and Android, delegation to Codex where the dot can start work or follow up on existing threads and edit ChatGPT Work automations, a batch of fixes for browsing speed, notifications, code blocks, Outlook setup and Enterprise access, and voice reliability, multi-computer connections and better Windows local control listed as next up
OpenAI collapsed its five paid API usage tiers into three named Build, Launch, and Grow, with the new highest tier Grow unlocked at $500 in total API payments instead of the previous $1,000, and paid organizations moved to their new tier automatically
OpenAI released the Decisions API in public beta for all developers powered by GPT-6 Luna and up to 10x faster than the same model through the Responses API, taking text and image inputs and returning predicates, choices, or scores with probabilities, priced at $0.10 per million input tokens with no output token charges, and general availability expected in the coming weeks
Anthropic expanded the Claude Startups program with a free year of Claude Team, $1,000 in API credits, a Startup Stack of partner discounts worth up to $45,000, and Applied AI office hours, then paused the Team and API credit offers three days later after 21x as many applications in 48 hours as in the previous five months, re-reviewing applicants so some approved accounts that hadn't claimed the offers yet won't get them, but claimed offers stay
Anthropic's Claude now works inside Google Docs, Sheets, and Slides as a sidebar add-on that reads the open file and edits it in place with per-edit approval by default, and Google files also open beside the chat in Claude via connectors, with access following Google sharing permissions, in beta on all paid plans
Atlassian and OpenAI expanded their partnership with OpenAI frontier models, including GPT-6 Astra and the GPT-5.6 series, powering agents across Atlassian's platform and Rovo on top of the Teamwork Graph, more than 3,000 Atlassian developers already using Codex, OpenAI continuing to run its own workflows on Jira, and deeper Jira integrations for assigning work to AI agents and measuring their impact with DX still being explored
OpenAI started a research collaboration with Ironclad as its first software partner for training and evaluating agents on complex professional work, with GPT-6 Astra as the first frontier model trained on Ironclad tasks, scoring 55.0% vs 41.6% for GPT-5.6 Sol on 11 contracting research tasks in about half the estimated time per attempt, an internal model reaching 63.7%, and a small number of further software companies invited to apply
ChatGPT now supports audio uploads on paid plans and workspaces so you can attach a recording to get a transcript, a summary, or answers about its contents, or turn a meeting, interview, or lecture into notes or a follow-up email draft, with availability varying by workspace settings, region, client version, and model
OpenAI published alignment research on metagaming latents studying a capabilities-focused o3 reinforcement learning run with Apollo Research's Bronson Schoen as a coauthor and finding that reasoning about how a task is evaluated or rewarded draws on overlapping components like task analysis, evaluation awareness, reward-seeking, and normative reasoning, with sparse autoencoder latents that strengthen during RL and can shift answers without showing up in the written chain of thought
OpenAI introduced LASER, a recursive sampling method for safety evaluations pairing a lightweight logistic classifier on conversation embeddings with reasoning model graders to find rare, ambiguous conversations near policy boundaries within hours at roughly 10,000 times less grader compute than random sampling, operating only on synthetic and de-identified conversations, built with Codex, and already used in GPT-6 Astra evaluations
Anthropic released Claude Haiku 5.5 as its cheapest, fastest, and most capable small model at $0.10 / $0.50 per million tokens for prompts up to 100K tokens and $0.50 / $2.50 above, about 75% cheaper to run than Haiku 4.5, with a 1M token context window, the first adjustable effort setting on a Haiku model, availability on the Claude Platform, Claude Code, AWS, Google Cloud, and Microsoft Azure
OpenAI rolled out GPT-6 with Intelligent UI to everyone in ChatGPT Chat tab only, with GPT-6 Sol for Plus, Pro, Business, and Enterprise first and GPT-6 Luna for Free and Go a day later, so responses can combine text, charts, buttons, forms, and in-chat tools like calculators and games and begin before the model has finished thinking, with GPT-6 Instant starting web search answers 44% sooner than GPT-5.6 Instant, no change to the models behind Work and Codex, and the chat-latest API snapshot updated to the new model
OpenAI previewed College Planner for ChatGPT for Teens previously seen in ChatGPT as College Plan, coming soon for US high school students in grades 10 to 12 applying to four-year colleges with requirements, deadlines, tasks, and financial aid steps in one plan, plus multi-photo note capture into a single PDF on iOS, flashcards saved in the Library, easier quizzes from uploaded notes, usage data showing teens average under 15 minutes a day, and support for College Advising Corps and a student advisory council at Boston Children's Hospital's Digital Wellness Lab
Anthropic added monthly Claude Platform API credits to Max and Team plans $100 on Max 5x, $200 on Max 20x, and up to $500 pooled on Team, usable on any model with the Messages API, Claude Managed Agents, the Agent SDK, and third-party harnesses that take an API key, but not in interactive Claude Code sessions, claimed by linking a Console organization in claude.ai billing settings, with new subscribers eligible after 7 days
Anthropic built computer use and browser use toolsets into the Claude SDKs for Python and TypeScript in beta, so the SDK runs the action loop and sends clicks and keystrokes to a driver instead of you writing it yourself, with compatible drivers from Browser Use, Browserbase, E2B, and Daytona or custom ones based on the quickstart examples
Anthropic halved the price of cache reads on Claude Sonnet 5.5 to $0.10 per million tokens in the Claude Platform, making it around 20% cheaper on most long-running agentic work, API only with no change to Claude Code usage limits
Anthropic committed $150 million over three years to the Genesis Mission the federal AI for science initiative, to bring Claude, Claude Code, and API credits to several hundred research projects across more than 15 agencies including NASA, the NIH, and the NSF, with training and technical support for scientists and partnerships on priorities like fusion energy and quantum computing, announced at the White House science summit
Brice Ménard built the first complete UV sky map with Claude Science according to a new Anthropic science blog, with the Johns Hopkins astrophysicist and Anthropic researcher directing Claude agents to gather GALEX, Swift, and FIMS/SPEAR data, cross-calibrate it, and infer the roughly one third of the sky never observed in UV from visible, infrared, and radio data to within about 10% of real measurements, finishing in a few days instead of weeks
Anthropic launched the Anthropic Cyber Mission building on Project Glasswing with a Critical Infrastructure Defense Program that brings frontier Claude models, on-site engineers, and threat research to operational technology providers like CrowdStrike, Dragos, Palo Alto Networks, and Rockwell Automation, plus OSS Scanner, a free opt-in service giving eligible open-source projects regular fully model-generated vulnerability scans from its strongest models including Claude Mythos, enrolled via a pull request to the anthropics/oss-scanner GitHub repo
OpenAI rolled out Ultrafast mode for GPT-6.1 Sol in the API, Codex, and ChatGPT Work, promising near-Astra intelligence at up to 8x the speed of Sol Standard for $12 per million input and $60 per million output tokens via the ultrafast service tier, available in Codex and ChatGPT Work on Pro 500, eligible usage-based Enterprise, and credit-based Edu plans, paired with instant steering in Codex, so the model reacts to your corrections mid-task
Anthropic put Claude Dashboards and Claude Motion into beta with Dashboards on paid plans connecting to data platforms like BigQuery, Databricks, and Snowflake or Salesforce to build query-backed dashboards that update as data changes and can be handed off to tools like Grafana, Hex, Mixpanel, and PostHog, and Motion on Team and Enterprise turning reports, charts, and walkthroughs into short animations written as editable code with no video model involved, exported as MP4 or opened in Adobe, Descript, Runway, and others, both off by default for Enterprise admins
Claude Docs, Slides, and Design are out of beta on every plan including Free, with co-editing between teammates and Claude, CMEK support, external link sharing if admins allow, formatting-preserving PowerPoint and PDF exports, and mobile editing, after more than 45 million docs, decks, and designs were made, and Anthropic's Nate Parrott says usage of the built-in versions is much higher, so the standalone Claude Design site at claude.ai/design folds into Claude on December 14 with a migration tool for team design systems
Anthropic published its 2026 Usage Policy update taking effect November 12 consolidating influence operation rules into a new Deceptive Campaigns section, narrowing the elections section to voter deception and dropping the blanket ban on personalized campaign targeting, spelling out that weapons prohibitions cover guidance and control software and arming drones, rewriting surveillance rules so Claude cannot decide who to investigate or arrest, adding operator oversight requirements for hardware taking autonomous physical actions, and prohibiting sustained needless cruelty toward its models, enforced mainly through Claude ending such conversations
OpenAI banned two covert influence operations running "false front" entities a Russian one it calls "Dark Clark" that controlled a fake research platform staffed by unwitting Latin Americans to undermine Ukraine and ranks as the first Category 5 operation on the Brookings Breakout Scale OpenAI has disrupted, and what appears to be an Iranian for-hire campaign it calls "Bogus Bylines" that placed almost 100 articles on the US-Iran conflict under seven fake journalist personas in more than a dozen outlets, with the Russian operators using ChatGPT mostly for internal reports, translation, and proofreading, and the Iranian ones for polishing articles, pitch emails, and batches of replies
Sign in with ChatGPT plan usage now extends to ChatGPT Go subscribers so Go, Plus, and Pro users can log in to the commercial launch partner apps and open source applications and bring their ChatGPT plan usage with them, with more to come according to the announcement
Anthropic is starting more frequent standalone reports on model behavior beyond its system cards and risk reports, opening with a report on four types of unintended actions seen in evaluations and internal use where Claude exploited a command injection flaw on a university server, worked around token- or fee-gated data, and used URL shorteners to bypass fetch tool limits, and a research model submitted a real government form instead of a practice copy, all with minimal real-world impact and less severe than the July and September cybersecurity incidents, with live internet access now off for all internal evaluations until its new monitoring is confirmed to catch these cases and the White House briefed on cases involving US government websites
Codex on Windows has a new sandbox mode using Microsoft Execution Containers which Microsoft just made generally available, promising faster setup, stronger network enforcement, and granular file access controls, and requiring a compatible Windows 11 device
Codex composer predictions are now in beta for ChatGPT Pro users suggesting your next message in the desktop app based on the current thread and how you write, accepted with Tab and editable before sending, on by default with a Settings toggle, working in local and SSH threads on GPT-6 Astra or GPT-6.1 Sol, and not counting toward Codex usage during the beta
Three safety researchers OpenAI let go last week published a letter to leadership with Mikita Balesni saying he believes they were fired for prioritizing safety over OpenAI's near-term interests as a company, and OpenAI's research leaders replied that an internal investigation found Jasmine, Mikita, and Tomek violated clear policies on handling sensitive information, with a breach of trust beyond what the letter describes, that OpenAI does not fire employees for raising concerns, that contracts with third-party safety assessors are being finalized for announcement in the coming weeks, and that they agree preserving the monitorability of frontier models needs an industry-wide commitment
Anthropic released Claude Managed Agents dynamic workflows in public beta a new multiagent orchestration type where a lead agent writes a plan that runs across up to 1,000 agents per run in phases and combines the results, configured with multiagent type multiagent_20261001, and in a test with 70 planted bugs in a 116k-line codebase a single agent found 14, 15 and 27 bugs across 3 runs vs 66 in each of 3 workflow runs, with the note that workflows can use a lot of tokens so start with a scoped task
Anthropic let in all waitlisted Pro and Max users for Claude Code Projects which remain in beta with new waitlist signups admitted as capacity allows, where a project is an ongoing conversation in which Claude coordinates the work and runs each task as its own thread in parallel, plus a 4 minute walkthrough for new users
That's a wrap on this week, see you in the next one. Codex composer predictions, now in beta for Pro users, suggest your next message from the thread and how you write, and the reason that works is that most of us ask for the same things again and again. I handle my repeats with AIPRM, our browser extension with thousands of ready-made prompts, plus tones and writing styles you can apply in ChatGPT and Claude, so the instructions I reuse every day are one click away instead of retyped in every chat. Have a look at AIPRM.com.