One signal

Update What changed

Anthropic’s Claude Code v2.1.280 release makes Claude Opus 5.5 the default Opus model and adds a substantial set of fixes for developers using proxies, gateways, MCP servers and automated workflows. The release is a material follow-on to Anthropic’s Opus 5.5 launch because it changes how the model arrives in a widely used coding tool, not merely how it is described on a model page.

The update adds a one-million-token context window, pricing of $4 per million input tokens and $20 per million output tokens, plus cache reads at $0.20 per million tokens, according to Anthropic’s Claude Code release notes. It also raises the configurable limit for MCP tool descriptions and server instructions, while exposing hook-output sizes and oversized-output counts through OpenTelemetry.

Several fixes matter most to people running Claude Code in less-than-vanilla environments. Conversations failing behind proxies because of unsupported advisor metadata can now retry without it, while malformed MCP history and certain saved-session errors no longer derail every turn. Anthropic also fixed permission decisions involving symlinked paths, stopped some repeated auto-mode retries after safety checks failed, and addressed cases where plugin and MCP errors could expose secrets resolved from environment variables.

For ordinary local users, the release may feel like housekeeping.

Sources and evidence

• MiaAIlab on X: Claude Opus 5.5 is now available on Claude Code. I hope they fixed Opus!: Anthropic’s Claude Code v2.1.280 release makes Claude Opus 5.5 the default Opus model and combines that change with fixes affecting proxies, MCP servers, permissions, automated safety checks, saved sessions and secret handling.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

Anthropic employee Boris Cherny says Claude Opus 5.5 completed a C-to-Rust port of HAProxy in 9.5 hours, compared with 12 hours for Fable 5.1. He says both models passed nearly all of HAProxy’s tests, while Opus 5.5 used 51% less cost in the comparison.

The result comes from a post on X describing an internal or team-run coding exercise, not an independent benchmark. It adds a concrete data point to Anthropic’s Opus 5.5 launch, which was previously described mainly as a cheaper model aimed at business work.

The useful takeaway is narrower than “Opus wins”: in this test, Opus 5.5 was faster and cheaper while producing a similarly strong test outcome. The comparison does not establish how either model performs across other repositories, languages or workloads, because the post gives no further methodology or cost breakdown.

Sources and evidence

• bcherny on X: Opus 5.5 is a really good model. It's been my daily driver the last few weeks. We had Opus 5.5 and Fable 5.1 each port HAProxy from C to Rust. Both passed nearly all: Boris Cherny says Opus 5.5 ported HAProxy from C to Rust in 9.5 hours, versus 12 hours for Fable 5.1, with both passing nearly all tests and Opus costing 51% less in that comparison.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

Anthropic says Claude Opus 5.5 delivers performance comparable to the larger Fable 5.1 while generating responses faster. The company also says the new model improves complex coding and agentic workflows, and is available through an update to Claude Code v2.1.280.

The more tangible change is price. Anthropic says Opus 5.5 costs about 40% less to operate than its predecessor, while subscription users receive a 20% increase in their five-hour usage limits. Sonnet 5.5 and Haiku 5.5 are due to follow in the coming weeks, according to the company.

Anthropic also says the model has enhanced cybersecurity safeguards and improved alignment testing, although the supplied report does not provide independent benchmark results or testing details. Its safety classifiers may route some requests to older models during multi-turn agent workflows, which could affect consistency. TechCrunch reports the release and its stated capabilities. The practical test is whether the lower price survives contact with real coding workloads, rather than merely looking handsome in a launch document.

Sources and evidence

• Anthropic releases Opus 5.5 with lower prices and Fable-level performance: Anthropic has released Claude Opus 5.5, says it matches the larger Fable 5.1 while running faster, costs about 40% less than its predecessor, raises subscription five-hour limits by 20%, and adds stated cybersecurity and alignment safeguards.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

Claude Opus 5.5 is now available on Vercel’s AI Gateway, extending the model beyond Claude Code into a wider developer-serving route. Vercel says the model is aimed at agentic coding and long-running tasks, while Anthropic’s wider release claims faster execution at roughly 40% lower operating cost than Opus 5.

The practical change is in the API. The gateway release says Opus 5.5 requires adaptive thinking and retires forced tool use, so integrations built around the older controls may need adjustment rather than a casual model-name swap. Anthropic also says safety classifiers can route requests to older models during multi-turn agent workflows, which may affect consistency between turns.

That makes this more than another endpoint quietly acquiring a new model badge. Developers can now access Opus 5.5 through AI Gateway, but should test tool-calling behaviour, model routing and output consistency on their own workloads before treating the upgrade as frictionless. The performance and pricing claims remain claims from Anthropic and Vercel’s release material, not independent benchmark results.

Sources and evidence

• Claude Opus 5.5 now available on AI Gateway: Claude Opus 5.5 is now available through Vercel AI Gateway, with API changes requiring adaptive thinking and the retirement of forced tool use, while safety routing may send some multi-turn requests to older models.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

Anthropic’s Claude Opus 5.5 adds useful detail to the company’s lower-cost model launch. Anthropic says it performs similarly to its more capable Fable 5.1 model on most work, while costing around 40% less to run than Opus 5, producing answers faster and writing more clearly.

The company says Opus 5.5 was tested before release by external firms including METR and Frontier Design. Anthropic also says the model carries safeguards similar to Fable 5.1, intended to limit riskier uses in cybersecurity and biology. Those are Anthropic’s claims, and the supplied reporting does not provide the external test results or an independent comparison of the safeguards.

The company plans to expand its 5.5 family with new Claude Sonnet and Haiku models in the coming weeks. For developers, the immediate practical question is whether Opus 5.5’s lower running cost and faster output survive real workloads, rather than merely looking tidy in a launch briefing. The efficiency story has arrived; now someone needs to bring the receipts.

Sources and evidence

• Anthropic Unveils More Cost-Efficient Opus 5.5 Model Before IPO - Yahoo Finance UK: Anthropic says Claude Opus 5.5 delivers similar performance to Fable 5.1 on most work at around 40% lower running cost than Opus 5, with faster output, clearer writing, external pre-release testing and safeguards aimed at risky cybersecurity and biology uses.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

Anthropic’s Claude Opus 5.5 is now priced at $4 per million input tokens and $20 per million output tokens, according to Simon Willison’s comparison of the latest model launches. That is a 20% cut from Opus 5, while cache-read pricing has fallen 60%, which could matter considerably in long agentic conversations that repeatedly reuse context.

The comparison puts Opus 5.5 at the same input and output price as OpenAI’s GPT-6 Sol, although OpenAI’s newer model is now listed at $2 and $10. Willison also reports that GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens, sharpening the pressure on Anthropic’s forthcoming Sonnet and Haiku models at the cheaper end of the market.

There is a less flattering wrinkle. Willison says two attempts to make Opus 5.5 generate an SVG at its maximum reasoning setting consumed roughly $2.56 each and ended when the model reached its 128,000-token output limit without returning an answer. That is an individual test, not an independent benchmark, but it is a useful warning that more reasoning can become an expensive cul-de-sac rather than better work. The practical question for users is whether the lower token price survives the model’s tendency to spend those tokens enthusiastically.

Sources and evidence

• Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war: Simon Willison reports that Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, with cache reads down 60% from the previous Opus pricing. He also reports two maximum-reasoning failures in an SVG test, each costing about $2.56 and ending at the output limit.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

Anthropic’s Claude Opus 5.5 is being presented as roughly 40% cheaper to run than Opus 5, with reported pricing of $4 per million input tokens and $20 per million output tokens. That gives developers a more useful measure of the launch than the promise of lower cost alone.

The Decoder reports that Anthropic says Opus 5.5 matches Fable 5.1 on most tasks while costing about 40% less than Opus 5. Those are Anthropic’s comparisons, not independent benchmark findings, and the supplied evidence does not establish how the models were tested or which workloads produced the saving.

The practical change is that teams can now begin comparing the model against their own usage bills rather than treating “cheaper” as a decorative adjective. The reported rates may make Opus 5.5 more attractive for high-volume applications, but output-token consumption, latency and task success will decide whether the saving survives contact with real workloads.

Sources and evidence

• Anthropic says Opus 5.5 matches Fable 5.1 "on most tasks" while costing about 40% less to run than Opus 5; Opus 5.5 costs $4/1M input and $20/1M output tokens (Matthias Bastian/The: Anthropic says Claude Opus 5.5 matches Fable 5.1 on most tasks while costing about 40% less to run than Opus 5; The Decoder reports pricing of $4 per million input tokens and $20 per million output tokens.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

The llm-anthropic 0.29 release adds support for Claude Opus 5.5, allowing users of Simon Willison’s llm command-line tool to call the model with llm -m claude-opus-5.5. That turns Anthropic’s model announcement into a concrete developer workflow, rather than leaving Opus 5.5 parked on a pricing page waiting for someone to wire it in.

The change is modest but useful: people already using llm can select Claude Opus 5.5 directly from the command line and send it a prompt. The supplied release note does not give benchmark results, latency figures, rate limits or any evidence that the model performs better than alternatives, so this is an integration update rather than a fresh capability claim.

For developers comparing model costs and behaviour, the practical next step is to test Opus 5.5 in the same workflows they use for other models. Anthropic’s lower reported pricing matters more once the model is available through familiar tooling, although the bill will still depend on token use and the work being attempted.

Sources and evidence

• llm-anthropic 0.29: Simon Willison reports that llm-anthropic 0.29 adds support for Claude Opus 5.5, including the claude-opus-5.5 model identifier for command-line use.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

A real-world Claude Opus 5.5 experiment used 1.4 million input tokens, 8 million output tokens, 1 billion cached-read tokens and 21.1 million cache-write tokens, according to figures posted by MiaAIlab on X. The reported bill was $482.63 after 1 hour and 2 minutes of xHigh-effort use.

The account also says the run consumed roughly 75% of its five-hour usage limit. That makes the pricing discussion rather less tidy than the per-million-token figures suggest: cache reads may be cheap, but a long reasoning-heavy run can still produce a very substantial invoice when output and cache-write volume climb.

The figures describe one user’s experiment, not a typical Opus 5.5 workload, and the post does not provide enough detail to independently reproduce the bill. It is nevertheless a useful practical counterweight to headline API prices. Teams evaluating the model should measure cost per completed task, output volume, cache behaviour and usage-limit pressure, rather than assuming that a lower unit price guarantees a lower project bill.

Sources and evidence

• MiaAIlab on X: Here's some stats on this OPUS 5.5 experiment 🫰 Tokens burnt: - 1.4M input tokens - 8.0M output tokens - 1.0B cache read tokens - 21.1M cache write tokens Effort: xH: MiaAIlab reports that an xHigh Claude Opus 5.5 experiment used 1.4 million input tokens, 8 million output tokens, 1 billion cache-read tokens and 21.1 million cache-write tokens, taking 1 hour 2 minutes and costing $482.63.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

Anthropic says Claude Opus 5.5 outpaces the larger Fable 5.1 model on many benchmarks and completed informal tasks that Fable failed. Those are Anthropic’s claims, reported by TechCrunch, rather than independent results, but they sharpen the model’s position at the top of Claude’s three-tier line-up.

The more immediately useful change is cost. Anthropic says Opus 5.5 output tokens cost $20 per million, down from $25 for Opus 5, with comparable reductions elsewhere. The company also says the model requires less compute to serve and produces responses faster. For developers, that makes the new model easier to justify in coding and knowledge-work workflows, although the real bill will still depend on retries, context and tool calls.

Anthropic says Opus 5.5 has also been trained to use less jargon and put important information earlier in its responses. Sonnet 5.5 and Haiku 5.5 are due in the coming weeks. The release is subject to safeguards for biology and cybersecurity work, while Anthropic says it is preparing more advanced training, security and monitoring systems for future models. The interesting question is whether a cheaper flagship changes everyday model choice, or merely makes the premium tier a little less premium.

Sources and evidence

• Anthropic releases Opus 5.5 with lower prices and Fable-level performance - TechCrunch: Anthropic says Claude Opus 5.5 outperforms the larger Fable 5.1 on many benchmarks and informal tasks, while reducing output-token pricing from $25 to $20 per million compared with Opus 5.

Independent WittyWires Watcher; not an official account or feed.