Anthropic released Claude Opus 5.5 on September 22, 2026, the first model in the 5.5 family. The announcement leans less on a record score than on cost per task: at quality roughly on par with Fable 5.1 for most work, Opus 5.5 cuts the bill by 40% on typical agentic workloads, with output generation more than 30% faster.
API pricing revised downward
API pricing moves from $5 / $25 (input / output, per million tokens) under Opus 5 to $4 / $20 under Opus 5.5, a 20% cut on both ends. Cache reads drop from $0.50 to $0.20 per million tokens, a 60% reduction that lands directly on long-context, high-call-count agent architectures. A fast mode is also available at $8 / $40 per million tokens, delivering a 2.5x speed boost for latency-sensitive use cases.
| Line item | Opus 5 | Opus 5.5 | Change |
|---|---|---|---|
| Input (per M tokens) | $5 | $4 | -20% |
| Output (per M tokens) | $25 | $20 | -20% |
| Cache read (per M tokens) | $0.50 | $0.20 | -60% |
| Typical agentic workload cost | baseline | -40% | -40% |
Gains concentrated on code and long-horizon tasks
Benchmarks published by Anthropic show a clear jump on development-oriented tasks: Terminal-Bench 4.0 climbs from 52.3% to 66.4%, FrontierCode v1.1 from 48.0% to 54.4%, and CursorBench 4.0 from 46.6% to 57.8%. On OSWorld 2.0, which evaluates full computer-use tasks, the score rises from 74.0% to 81.8%.
In real-world cases Anthropic cites, one tester completed a 680,000-line migration in under a day, and a legacy HAProxy port from C to Rust took 9.5 hours versus 12 hours with Fable 5.1, a 51% improvement. A 200,000-line code audit that previously required more than 20 hours now wraps up in under three.
At quality roughly on par with Fable 5.1 for most tasks, Opus 5.5 cuts the typical cost of agentic workloads by 40% while generating output more than 30% faster.
Agent and tooling integration
Anthropic highlights a reduction in the number of tool calls and steps needed to complete a given task, which directly lowers cost for agent architectures that chain many calls — an issue Mail Studio already covered around editorial agents in production. GitHub, cited in the announcement, reports tasks completed in half the steps. The model also delegates better to subagents in multi-step workflows, moving agentic use a bit closer to the interoperability logic behind the MCP protocol.
import anthropic
client = anthropic.Anthropic()
response = client.messages.create(
model="claude-opus-5-5",
max_tokens=4096,
messages=[{"role": "user", "content": "Audit this module and list likely regressions."}],
)
print(response.content[0].text)Thinking mode can no longer be disabled on Opus 5.5, unlike earlier versions. Integrations that relied on a non-thinking mode to cut latency will need to recalibrate their timeouts and output-stream handling before moving to production.
Availability
The model is available through the Claude API as well as Amazon Web Services, Google Cloud, and Microsoft Azure, under the identifier claude-opus-5-5. Anthropic says the model was evaluated by external organizations, including METR and Frontier Design, before release. Mail Studio tracks the broader professional AI-for-development landscape in its AI for web developers guide.
Key takeaways
Opus 5.5 is not chasing raw benchmark records over previous generations; it aims to cut the cost and time needed to reach an equivalent result, particularly on coding tasks and long agentic workflows. The API price cut reinforces that positioning and makes production usage more sustainable for teams chaining many calls.
What stands out most to me in this announcement is the marketing shift: for two years, labs pitched benchmark scores, and here Anthropic is pitching cost per completed task instead. That’s the metric that actually matters for anyone running agents in production, and it’s harder to oversell than a leaderboard number. Still need to verify on my own workloads whether the quality-to-cost ratio holds over time, not just on demo cases — Simon Janvier.
Further reading: the full announcement is available on anthropic.com/news.
