Anthropic's Opus 5.5 lands with 40% lower costs and faster coding

If you have tried using AI coding agents on massive repositories lately, you know the drill: the session runs for hours, context explodes, and your token bill goes through the roof.

Anthropic just launched Claude Opus 5.5, kicking off its brand new 5.5 family. Arriving just two months after Opus 5 dropped in July, the new flagship cuts typical operating costs by 40% while generating output over 30% faster.

It also takes direct aim at developer expenses. Input and output tokens now cost $4 and $20 per million respectively, both down 20%. But the real headline for heavy terminal users is cached reads: Anthropic slashed that price by 60%, down to just $0.20 per million tokens.

Built for marathon coding sessions

Why obsess over cache pricing? Anthropic analyzed Claude Code telemetry from March to September 2026 and found a massive shift in user behavior. Developers now let Claude work 3.3x longer per prompt, request context grew 2.6x, and the input-to-output token ratio stretched to a staggering 324:1.

In plain English: modern agentic coding is all about reading huge amounts of code again and again, just to modify a handful of files. Lowering cache reads to twenty cents directly attacks the heaviest part of that bill.

In real-world tasks, one early tester completed a 680,000-line code migration in less than a day, a job that would normally tie up an entire engineering squad for weeks. In an internal benchmark rewriting HAProxy from C into Rust, Opus 5.5 crossed the finish line in 9.5 hours compared to 12 hours for Fable 5.1, cutting total project cost by 51%.

Benchmark scores followed suit. Opus 5.5 scored 66.4% on Terminal-Bench 4.0 and 54.4% on FrontierCode, outpacing OpenAI's GPT-6 Astra and GPT-5.6 Sol. Early tester Walleye Capital noted that the model even caught a logic flaw inside their evaluation instructions and corrected it on the fly, something no prior model had managed.

Tighter guardrails and cloud rollout

As capabilities climb, the safety measures are tightening as well. This marks Anthropic's first release since CEO Dario Amodei called to pace frontier advancement so safety practices can keep up.

Because Opus 5.5 rivals the unreleased Mythos 5.1 in biology and cyber capabilities, it ships with Fable-level safeguards. High-risk requests will trigger refusals unless organizations are vetted through dedicated verification programs.

The model is already available on Amazon Bedrock and the Claude Platform, alongside a new Fast mode delivering up to 2.5x speed at $8 per million input and $40 per million output tokens. Sonnet 5.5 and Haiku 5.5 are slated to follow in the coming weeks.

Why it matters

For developers and engineering leads, Opus 5.5 removes the biggest friction in agentic development: the compounding cost of long context. Running multi-hour autonomous refactoring on large codebases used to be an expensive gamble; slicing cache costs by 60% makes deep autonomous work genuinely affordable. If you are already building with Claude Code, switching now is an easy call.

Across the industry, Anthropic is playing a very smart game. Instead of simply racing for raw compute records, they tailored both their pricing and inference engine around how developers actually interact with agents. When real-world coding requires re-reading hundreds of thousands of tokens per turn, winning the cache price war is how you lock in developer loyalty.

This puts serious pressure on rivals like OpenAI to adjust their agent economics. But we think the aggressive safety interventions remain a potential friction point: Zapier's workflow evaluations saw lower scores simply because guardrails intervened without fallbacks. If enterprise agents get tripped up by overly eager classifiers, teams will still need to keep secondary models in their pipelines.