Anthropic launched Claude Opus 5.5 on September 22, 2026, billing it as the first model in the new Claude 5.5 family. The headline numbers: input and output tokens are priced at $4 and $20 per million respectively, 20% below Opus 5, and cache reads drop to $0.20 per million tokens, 60% below what Opus 5 charges. Anthropic says the model generates output more than 30% faster than Opus 5 and costs 40% less to run on typical workloads.
That last figure is the one worth unpacking. A 20% cut in token pricing and a 40% reduction in overall workload cost are different numbers. Cache reads, Anthropic says, “make up the majority of agentic and coding work costs.” At $0.20 per million tokens versus the old $0.50, the cache read price is doing the heavy lifting. For applications that process the same large context over and over, like agents that keep a codebase or document in the window, that 60% reduction in cache reads compounds into the bigger number fast.
If you build systems where most tokens are cache hits, Opus 5.5 is materially cheaper than the raw per-token prices suggest.
What Anthropic Is Claiming
Anthropic positions Opus 5.5 as performing at the level of its Claude Fable 5.1 model on most work, which puts it at the top of its publicly available lineup for coding and knowledge tasks. On Terminal-Bench 4.0, an agentic coding benchmark, Anthropic reports a score of 66.4% for Opus 5.5 versus 55.8% for Fable 5.1 and 52.3% for Opus 5.
The company offers a few concrete examples. One early tester, Anthropic says, completed a 680,000-line code migration in under a day. On an internal web app optimization test, Opus 5.5 succeeded in cutting load times 39 of 40 times while Opus 5 made smaller changes that also altered app behavior (Source). These are Anthropic-sourced examples, not independent benchmarks, so the usual caveats apply. They describe what the model can do under favorable conditions, not what it will do on your specific workload.
The model is available now on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS, using the model ID claude-opus-5-5. Anthropic says Claude Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks.
The Pricing in Context
For teams comparing Opus 5.5 against what they could get from OpenAI or Google, the picture is more complicated.
OpenAI’s current top-of-line model, GPT-6 Astra, runs $10 per million input tokens and $50 per million output tokens on its standard API. Opus 5.5 at $4 input and $20 output is 60% cheaper on both. For shops running high-capability inference at volume, that gap is real money.
The more direct OpenAI coding comparison is GPT-6 Sol, which OpenAI describes as built for coding and agentic workflows. That model runs $2 per million input tokens and $10 per million output. Cheaper than Opus 5.5 on both dimensions. Cached input on GPT-6 Sol is $0.20 per million, identical to Opus 5.5’s cache read price.
Google’s current flagship Pro model, Gemini 3.1 Pro Preview, prices at $2 per million input tokens and $12 per million output tokens for prompts under 200,000 tokens. Also cheaper than Opus 5.5 on raw input and output.
So Opus 5.5 is not the cheapest capable model on the market. What Anthropic is selling is a combination of high benchmark scores and lower cost than its own previous flagship, not a price floor across the industry. The argument for Opus 5.5 over GPT-6 Sol or Gemini 3.1 Pro rests on capability, not price. Teams considering switching from a competitor face a typical build-versus-benchmark decision.
For very high-volume caching-heavy workloads, the $0.20 cache read rate is a useful data point. At that price, Opus 5.5 ties GPT-6 Sol on cache reads.
Who This Actually Matters To
The pricing change is most consequential for teams already running Opus 5 at scale. Anyone hitting the API cost ceilings that enterprise teams have run into with Anthropic’s flagship should update their model string.
For agent builders specifically, the cache read discount is structural. Agents that hold long system prompts or large codebases in context call the model dozens or hundreds of times per task, and most of those calls are primarily cache hits. The cost model for long-running autonomous workflows changes meaningfully when cache reads are this cheap.
Teams evaluating a first-time Anthropic adoption should treat the 40% workload cost figure as optimistic. That number reflects Anthropic’s own test conditions, which presumably favor caching-heavy patterns.
Safety and What Comes Next
Anthropic says Opus 5.5 scores highest on its internal behavioral audit, a test it describes as covering thousands of simulated scenarios. The company says the model is less likely than recent releases to take hard-to-reverse actions or operate outside given boundaries. Given the model’s positioning for long-running autonomous tasks, those properties matter to anyone thinking about deploying it unsupervised.
Anthropic also notes it is deploying the model with safeguards similar to those on Claude Fable 5.1 due to comparable capability levels in biology and cybersecurity domains.
Batch API pricing for Opus 5.5 is $2 per million input tokens and $10 per million output, the standard 50% discount applied to async workloads. Fast mode, still in research preview, prices at $8 per million input and $40 per million output for faster throughput at premium cost.
The rest of the 5.5 family is still pending. Anthropic says Sonnet 5.5 and Haiku 5.5 follow in the coming weeks. If those carry similar relative cost improvements, the pricing dynamics across the full Claude lineup shift again.
Source: Anthropic announcement (September 22, 2026); Anthropic pricing documentation; reporting by Emergent.