GPT-5.6 ships, Fable fights back, and Claude Code gets a browser

OpenAI's GPT-5.6 lands as the default in M365 Copilot, Anthropic extends Fable 5 access under pricing pressure, and Claude Code gains browser control.

The big picture

OpenAI had a banner Friday, dropping GPT-5.6, making it the default inside Microsoft 365 Copilot, and watching a variant reportedly crack a 50-year-old math problem. The downstream pressure on Anthropic is visible in real time: Fable 5 keeps getting its sunset date pushed back as the pricing war heats up. Meanwhile, the tooling layer is quietly getting more capable — Claude Code now has a built-in browser, vLLM got a native transformers backend, and structured agent memory turned in a surprisingly strong result in a game environment.

GPT-5.6 arrives with real momentum behind it

OpenAI released GPT-5.6 today, framing it as a model that delivers more intelligence per token at better price-to-performance than its predecessors. The headline positioning is “frontier intelligence that scales with your ambition” — which is marketing — but the substance is that this is the model now powering Microsoft 365 Copilot as its default across Word, Excel, PowerPoint, Chat, and Cowork. That’s not a beta slot; it’s the model your colleagues are hitting every time they click the Copilot button. OpenAI / OpenAI M365 post

On the capability end, the more striking story comes from GPT-5.6 Sol Ultra, a variant that reportedly produced a proof of the Cycle Double Cover Conjecture — an open graph theory problem that has resisted mathematicians for roughly 50 years — in under an hour using 64 parallel subagents. Mathematician Thomas Bloom reviewed the proof and called it surprisingly elementary, though he noted it lacks citations for prior work that informed it. The honest question here is whether this constitutes genuine mathematical discovery or very sophisticated pattern completion across a corpus that almost certainly contains all the relevant lemmas. That debate is going to run for a while. What’s less debatable: 64-agent parallel reasoning is a different kind of tool than a single-turn chat assistant, and it’s starting to show. The Decoder

OpenAI also published a pointed analysis of SWE-Bench Pro, one of the most widely cited coding benchmarks in the industry. Their review raises concerns about reliability and signal quality in how the benchmark evaluates AI code performance — specifically whether scores reflect genuine capability or artifacts of how the benchmark is constructed. This matters a lot: if the benchmark is noisy, leaderboard comparisons between models become harder to trust, and it makes the “we scored X on SWE-Bench” press releases from every lab worth scrutinizing more carefully. OpenAI

Anthropic holds the line on Fable 5 as pressure mounts

Anthropics Claude Fable 5 was supposed to drop off paid plans and move to pay-per-use today. Instead, Anthropic extended access through July 19, for the second time, letting subscribers burn up to half their weekly usage limit on Fable 5 and keeping Claude Code rate limits 50% higher than normal. The reason isn’t hard to find: GPT-5.6 Sol is available without usage restrictions on Plus, Business, and Pro plans right now, and OpenAI just hit 6 million active users on its Codex and ChatGPT Work products. The Decoder / Simon Willison

Simon Willison makes the case plainly: Anthropic should just keep Fable permanently on those plans. The rolling deadline creates uncertainty that actively drives users toward OpenAI, where you know what you’re getting. From a product strategy standpoint, repeated one-week extensions are the worst of both worlds — you neither capture revenue nor retain subscriber confidence. Anthropic is in a tough spot computationally, but the messaging is making it worse.

Separately, Claude Code shipped a built-in browser that lets the agent open, navigate, read, and interact with external websites directly inside the development environment. Write actions on external sites — form submissions, purchases, account creations — are gated behind classifiers and require user approval before executing. This is a significant capability jump: rather than copy-pasting content from the web into a chat, the agent can now check documentation, look up API specs, and verify its own output against live systems in one loop. The approval gates for write actions are the right call given prompt injection risks from external sites. The Decoder

The open-source inference stack tightens up

Hugging Face published a post on a new native vLLM transformers modeling backend that runs at native vLLM speed. The significance here is that the Transformers library and vLLM have historically been two separate stacks — you’d prototype with Transformers and deploy with vLLM, often maintaining two code paths. A native backend that runs at vLLM’s inference speeds through the Transformers interface collapses that gap, which is genuinely useful for teams who want production throughput without the operational overhead of maintaining a separate serving framework. Worth watching for benchmarks from the community before fully committing. Hugging Face

NVIDIA and Hugging Face jointly published an open data initiative focused specifically on agentic use cases. The framing is that most public datasets were built for single-turn Q&A or instruction-following, not for training agents that need to plan, use tools, and recover from errors across multiple steps. Quality training data for agents is a real bottleneck right now — most agentic benchmarks have only hundreds to low thousands of examples — so a push toward open, purpose-built agentic datasets is the right problem to attack, even if the actual dataset contents weren’t detailed in what was available here. Hugging Face

Agent research: structured memory beats long context

A research project called AgenticSTS tackled one of the most frustrating practical problems in building LLM agents: context windows that balloon out of control. They replaced the standard approach — appending every interaction to a growing chat log — with five separate, purpose-specific memory layers. Tested on the card game Slay the Spire 2, which requires long-horizon planning and strategy updates, the structured approach kept prompt size around 5,000 tokens compared to unstructured logs that explode past 500,000. The agent won 6 out of 10 games; competing agents won zero. The Decoder

The token number is the headline. 100x smaller prompts means 100x cheaper inference, faster responses, and much less sensitivity to the “lost in the middle” problem where models degrade on information buried in long contexts. The game domain is narrow but the technique is general — this is the kind of memory architecture pattern that production agent builders should be reading. The challenge is that designing five coherent memory layers requires knowing what your agent needs to remember and forget, which is non-trivial for general-purpose agents.

Simon Willison also published a short but sharp piece arguing that LLM agents should never be considered a “Directly Responsible Individual” (DRI — the person ultimately accountable for a project’s success or failure, a concept originally from Apple). His point is simple: accountability requires the capacity to be held responsible, and a machine cannot be. It’s a useful framing for teams currently debating how much autonomous authority to give agents in their systems. Simon Willison

Enterprise AI deployments and the usage reality check

OpenAI published a case study on Deutsche Telekom becoming an “AI-native” telco — transforming customer service interactions, employee workflows, network operations, and voice interfaces using OpenAI’s models. The specifics are light (it’s a marketing case study), but the pattern is notable: telcos are a good bellwether for large-enterprise AI adoption because they have enormous customer contact volume and measurable cost structures. If Deutsche Telekom can quantify call deflection rates and handling time reductions, those numbers will travel fast through the rest of the industry. OpenAI

Anthropics analysis of 1.2 million Claude Cowork sessions across more than 600,000 organizations is more interesting than a case study because it shows actual usage distribution. About half of all sessions involve what Anthropic calls “work around the work” — status reports, onboarding checklists, slide decks, meeting summaries. Software development barely registers in Cowork because developers are on Claude Code. The takeaway for product builders: enterprise AI adoption is being led not by developers or power users, but by the administrative middle layer of every organization. Build for that if you’re targeting enterprise. The Decoder

S&P Global downgraded Oracle’s credit rating to BBB-minus — one notch above junk — citing OpenAI as a key credit risk. OpenAI represents roughly half of Oracle’s $638 billion in total contractual obligations. If OpenAI were to walk away, Oracle would be holding a mountain of data center capacity with no obvious replacement tenant. This is a genuinely unusual risk profile: a single AI company’s financial health is now material to the credit rating of one of the world’s largest enterprise software companies. The Decoder

AI, society, and the uncomfortable data points

A Brown University economics professor ran a natural experiment worth paying attention to. A take-home exam averaged 96 percent; he suspected widespread AI use and switched the final to an in-person, proctored format. Eighteen students dropped the course before the exam, nine more didn’t show up on the day, and the class average dropped to 48.6 percent. Two independent large-scale studies from China and UC Berkeley found the same pattern: students who lean on AI for assignments consistently underperform on proctored tests. For developers building educational tools or thinking about AI-assisted learning, this is a real feedback loop problem worth designing around. The Decoder

Sam Altman told an interviewer he is “pretty sure” AI has created more jobs than it has eliminated, which is a notable reversal from earlier warnings about entire professions disappearing. Anthropic’s Dario Amodei is walking back similar doomsday predictions. The empirical picture is genuinely unclear — studies so far don’t firmly support either extreme — but the optics of CEOs hedging their own prior predictions should inform how you weight their future forecasting. The Decoder

Meta pulled the Muse Image feature that let users generate AI photos of any public Instagram account holder by @-mentioning them, no consent required. Meta’s own statement was that the feature “missed the mark.” The feature lasted only days after launch, which suggests either that the consent problem wasn’t caught in internal review or it was caught and launched anyway. Neither is a good look, and the incident is a concrete example of why “you need approval to write” logic in tools like Claude Code’s browser matters. The Decoder

A Pangram study found that 1 in 4 longer posts across five social platforms is fully AI-generated, with LinkedIn leading at 41 percent of long-form posts flagged as AI-written — despite making up only a third of the total content scanned, LinkedIn accounted for nearly two-thirds of detected AI content. The detection model is described as conservative, so the real rate is likely higher. If you use LinkedIn for any kind of signal about your industry, adjust accordingly. The Decoder

Quick hits

Google’s SensorFM, a foundation model trained on a trillion-plus minutes of Fitbit and Pixel Watch data, beats benchmarks on 34 of 35 health tasks — no product integration announced yet. The Decoder
Waze is adding Gemini-powered conversational voice commands for traffic reporting and destination search, plus three other updates. The Verge
Apple’s cancelled self-driving car program apparently seeded the Neural Engine that now powers on-device AI in iPhones — a nice origin story, if you missed it. The Verge
OpenAI launched a Bio Bug Bounty program focused on GPT-5.5’s biosecurity properties, inviting external researchers to find failure modes. OpenAI
OpenAI published its principles for government and national security partnerships, covering democratic accountability and responsible use. Vague but worth a skim if you work in that space. OpenAI
OpenAI is hiring a dedicated product manager for family, caregiver, and older adult experiences as ChatGPT pushes deeper into household use cases. TechCrunch
Community opposition to AI data center buildouts is growing, with local fights over power grid impact tracing back to a 2015 Apple data center dispute in Ireland. The Verge
Simon Willison released sqlite-utils 4.1 and 4.1.1 in quick succession — the former adds a --code option for insert/upsert, the latter patches a foreign key / transaction edge case that Claude itself spotted. Simon Willison
shot-scraper 1.11 ships with better server startup polling (up to 30 seconds instead of a fixed 1-second delay) and a --js-file option across multiple commands. Simon Willison