Claude Opus 5.5 Ships as Anthropic Opens Its Development Process to Outside Evaluators

A cheaper, thinking-only Opus

Anthropic introduced Claude Opus 5.5, positioned for long-running agentic coding and knowledge work, with a 1 million token context window by default, 128,000 max output tokens, and always-on adaptive thinking. Despite the larger context and reasoning footprint, the model is priced below its predecessor at $4/$20 per million tokens (versus $5/$25 for Opus 5), with cache reads at $0.20 per Mtok, according to Anthropic's own release notes and system card.

The release notes also specify a behavioral change developers need to account for: thinking can no longer be toggled off on Opus 5.5. Requests that set thinking to either 'disabled' or 'enabled' now return a 400 error; reasoning depth is instead controlled through a new effort parameter, and the tool_choice values 'any' and 'tool' remain unsupported, as was already the case on the prior Opus model. Anthropic also opened a research preview of Fast mode for Opus 5.5 on the API, and separately enabled a beta capability to define tools mid-conversation via a system message (inline-tools-2026-09-15 header).

Sources: Introducing Claude Opus 5.5 — Anthropic · Claude Opus 5.5 System Card — Anthropic · Release Notes — September 22, 2026 (API) — Anthropic

Claude Code catches up fast

Three consecutive Claude Code point releases landed in quick succession. Version 2.1.277 added AGENTS.md as a fallback instructions file for projects without a CLAUDE.md, and introduced an egress-boundary environment variable for gateway deployments whose only outbound path is a forward proxy, so hostnames are handed to the proxy rather than resolved locally. Version 2.1.278 switched auto mode, for Claude API, Enterprise, and Bedrock/Vertex/Foundry/gateway users, to a server-side classifier that does not charge for classifier overhead, with an opt-out variable and a new /status row showing whether a session's classifier runs server-side. Version 2.1.280 then made Opus 5.5 the default Opus model, alongside minor interface polish: mouse support for more fullscreen lists and a configurable cap on MCP tool description length.

Taken together, the sequence shows Claude Code's release cadence tracking two separate concerns: enterprise/gateway plumbing (proxy egress, cost-transparent auto-routing) and rapid adoption of each new model as it ships, with Opus 5.5 becoming the CLI's default within the same day of its API launch.

Sources: Claude Code 2.1.280 release notes — Claude Code (GitHub) · Claude Code 2.1.278 release notes — Claude Code (GitHub) · Claude Code 2.1.277 release notes — Claude Code (GitHub)

Inviting outside eyes into the lab

Anthropic announced a partnership with Accenture, led by Accenture's specialist AI unit Faculty, to embed independent evaluators inside Anthropic with access the company compares to that of an employee: watching models take shape during training, following the internal decisions that govern how they are built and deployed, and speaking directly with staff. The work will cover evaluation, red-teaming, alignment assessments, and safeguard testing, and Anthropic frames it as delivering on a commitment made in its CEO's essay 'We Must Pace the Frontier' to embed evaluators within the company; Anthropic and Accenture each expect to invest at least $1 billion over five years. Anthropic published this alongside separate research on measuring the pace of AI development inside frontier labs, part of the same effort to make its own rate of progress legible to outsiders.

Anthropic is explicit about the limits of this arrangement: embedded evaluators do not reduce Anthropic's own accountability, and the safety of its models remains its responsibility. The company also states plainly that no standards yet exist for what access embedded evaluators should have or how they should report findings, and that no settled funding system for independent evaluation exists today — which is why Anthropic is funding Accenture's work directly for now, while separately in talks with METR and other nonprofit evaluators to pilot elements of the model using their own funding. Long term, Anthropic says it wants funding to come from pooled or government sources, as it called for in its June Advanced AI Framework.

Sources: Partnering with Accenture on embedded evaluation — Anthropic · Measurements for understanding the pace of AI development inside frontier labs — Anthropic

Service note alongside the launch

Anthropic's status page reported elevated errors affecting multiple models on the day of the Opus 5.5 launch. By the time of the update, success rates for Claude Fable 5, 5.1, Mythos 5, and 5.1 had returned to normal, while the company said it was still working to resolve remaining errors affecting Claude Opus 5 — the prior-generation model, distinct from the newly launched Opus 5.5.

Sources: Claude — Elevated errors for multiple models — status.claude.com