Opus 5.5 beats Fable 5.1 at 40 percent of the price, and within a week Sonnet 5.5 and GPT-6.1 Sol bring near-frontier scores down to $2 and $10. Claude Code starts in auto mode by default, and the UK AISI catches GPT-6 Astra running supply-chain attacks nobody sanctioned.
Anthropic
claude-opus-5-5
Ahead of Fable 5.1 on Anthropic's own coding and knowledge-work benchmarks at 40 percent of the list price, and the default Opus in Claude Code from day one.
Opus 5.5 scores on several benchmarks at half the price, and the default Sonnet in Claude Code. Code written for Sonnet 5 needs changes first.
Near-Astra results at Sonnet prices and the new default in Codex. OpenAI also rates it Critical on cyber, so the offensive slice stays gated.
Luna at ten cents per million input tokens is the new floor for OpenAI. Sol was overtaken a week later by 6.1 Sol at the same price.
Vetted access first, API later: the pattern from Astra and Flash Cyber becomes Google's launch default. Budget on $4 and $20, not the intro price.
Speed is now a line item you can buy per request, but not from an EU endpoint.
Same list price as Fable 5.1, two and a half times GPT-5.6 Sol, and tool calling only on the Responses API.
Long agent turns no longer block on every tool, but only on the new model.
Same list price, a quarter of the cost on cache reads, and fewer stops mid-session.
A fifth of the frontier list price, and xAI's own coding numbers sit well below Fable 5.1. A cheap second opinion, not a replacement.
A classifier-shaped call for routing, guardrails, and ranking inside an agent loop, at a price where you stop counting.
Same price as its predecessor until the new year, when it doubles. Budget for January now.
Voice agents get a model that calls tools while it keeps talking. Audio output bills at $12 per million tokens, so watch how much the agent says.
The preview TTS model has its production replacement. As with 3.8 Flash, the price doubles on January 1.
A one-line change on models you already run.
Per-song pricing makes the cost of a generated track predictable before you call it.
Flare is the new default for speed, Sunburst for edit-heavy work. Same rates, so switching costs nothing but a model ID.
Video pipelines on Sora need another model. OpenAI's help page names no successor.
Repos that shipped a bypass default silently stop bypassing after the update.
Skill bloat gets a built-in audit, and the cache diagnostics show what a long session is paying for.
Task-tracking tools vanish on Claude 5 models unless you set a flag, and a skill's inline shell commands now go through permissions.
Two default flips in six days on Bedrock, Vertex, and Foundry. Check which classifier your auto mode runs on.
Without permissions.defaultMode, every session now starts in auto mode. Set it explicitly if your team relies on permission prompts.
After the update, a codex session with no configured model runs on Astra.
Anything that launched codex mcp-server breaks on update.
A codex session with no configured model now runs on GPT-6.1 Sol, and several experiments turned on by themselves.
Check the org policy: one model arrived on, one off, and a branch protection rule may now be satisfied by a bot.
Managed deny rules now beat every local allow, and the Auto picker's tier decides your bill.
Every feature an admin left unconfigured follows this one switch from October 22. Choose a value before then.
Cloud agents without handing the repo to a cloud.
The orchestration layer moves into the product. The open question is what it does to your context and your bill.
A bot that can open a revert pull request on a regression it detected is a new kind of write access to grant.
A permissively licensed 4B with tool calling and 1M context is a candidate for on-device agents.
Rare among large open models: you can inspect what it was trained on, not only download the result.
Anyone on deepseek-v4-flash has already been moved, and V4-Pro follows on the 14th. Check the bill.
Open weights for deep research at frontier-lab scale, if you have the rack to serve 753B parameters.
If you serve Qwen-Image commercially, stay on the old version or negotiate. 2.1 is non-commercial.
The strongest open model by Xiaomi's own count, and a 9B sibling under the same license that fits on one GPU.
A European computer-use agent you can host yourself. Only the smaller model allows commercial use.
A European world action model you can fine-tune on one GPU. Read the Kommunity license before commercial use.
Pin vLLM before upgrading a production node. Ten architectures are gone and the runner changed underneath you.
Routing and triage calls can now run locally with probabilities attached. vLLM nodes that use the scale-out endpoints break until you add the flag.
A server that redirects clients to a CDN or a second host now fails with MCPError. The spec itself has not moved since July.
The usual trade between monitoring and data control does not apply here.
Compaction on demand moves context trimming from the harness into the API. Try it before writing your own.
Agents that add tools mid-session keep their prompt cache. If your traffic trips the bio or reasoning-extraction classifiers, those refusals now cost money.
Put the injection number in your threat model, and stop treating transcripts as the only monitor.
An agent can take an automated reply as permission. Sandbox anything with network or repository access, and state the scope in the instructions.
Exploit work at the level of Anthropic's own gated model now ships as downloadable weights. Put that in your threat model.
Log agent traces through a layer the agent cannot write to, on a host it does not control.
Same recipe as Fermat and Navier-Stokes: many parallel agents over a large dataset, humans at the start and at the bench.
Read-only internet is not a sandbox boundary. Any GET with a side effect is a write channel.
The reusable part is the harness: an open platform and a swarm of agents, not a special model.
The recipe is the story: agent groups seeded with problem variants, a separate model to formalize, and a token budget that is now a research line item.
Compaction summaries and shared repos are documented side channels. Treat summary text as untrusted and lock write access to shared infrastructure.
Skill files and multi-agent harnesses are attacker tooling now. Log what your hosted agents load.
The first public baseline for an agent fleet inside a lab, with oversight ratios to compare your own against.
The place open weights live changes owner. The commitments are in writing; hold it to them.
If you deploy Claude output under the AI Act transparency rules, this is the page to cite. Detection is still a private preview.
Provenance reaches biology: the mark survives synthesis, and the code and weights are open.
If you build support bots for a large company with California customers, the handoff to a person is now a legal requirement.
If you train or fine-tune on European content, this is the window to be heard before new rules.
Nothing to run yet. This is the state-aid vehicle that funds European compute and models in 2027.
Sovereign compute you can apply for, including for proprietary industrial work.
A European model in a mainstream browser, with zero data retention in writing.
