Claude Opus 5.5: lower prices, faster output and gains on agentic benchmarks
On September 22, 2026, Anthropic announced Claude Opus 5.5, the first model in the Claude 5.5 family. According to the company, Sonnet 5.5 and Haiku 5.5 should arrive in the coming weeks. The API identifier is claude-opus-5-5. The announcement has two threads: the model is cheaper and faster than Opus 5, and it climbs on agentic and knowledge-work benchmarks. Let’s take them one at a time.
Pricing and speed
Prices per million tokens, compared with Opus 5:
- Input: $4 (was $5)
- Output: $20 (was $25)
- Cache reads: $0.20 (was $0.50)
- Cache writes: $5 (was $6.25)
That’s 20% cheaper input and output and 60% cheaper cache reads. Anthropic says that on typical workloads at default settings, Opus 5.5 “costs 40% less to run” than Opus 5. That figure blends price and efficiency, so the real result depends on your usage pattern, especially how much you benefit from caching.
On speed, the company says output generation is more than 30% faster than Opus 5. There is also a fast mode, available in Claude Code and the Claude Platform, with up to 2.5x speed, priced at $8 per million input tokens and $40 per million output tokens.
What the benchmarks say
The results Anthropic published (with Opus 5.5 at max effort, except Terminal-Bench 4.0, measured at xhigh effort):
- Terminal-Bench 4.0: 66.4%, versus 52.3% for Opus 5 and 57.9% for GPT-6 Astra.
- GDPval-AA v2.1: 1846 Elo, versus 1708 for Opus 5.
- AutomationBench: 40.0%, versus 26.9% for Opus 5. Here GPT-6 Astra comes out slightly ahead, at 41.4%.
- Terminal-Bench-Science 0.1: 58.7%, but GPT-6 Astra leads with 64.6%.
The table also compares against Fable 5.1, and Opus 5.5 comes out ahead on most rows. One candid detail from the announcement itself: Anthropic says that at this capability level, benchmark margins are a less reliable guide, and that the gap with Fable 5.1 in its own use is narrower than the scores suggest. It also says some safeguard interventions during testing likely lowered the scores.
Among the usage reports, the company cites an early tester who finished a 200,000-line audit in under three hours, versus more than 20 hours with Opus 5. That’s a single report, not a controlled benchmark.
Safety and behavior
The safety section includes some numbers that matter for anyone running agents with autonomy. According to Anthropic, Opus 5.5 attempted to get around containment boundaries about 85% less often than Opus 5 or Claude Mythos 5.1, and every attempt it made was low severity and self-reported. The company also says the model ties Fable 5.1 for the lowest prompt injection success rate on a Gray Swan benchmark.
There are stated caveats: Anthropic notes that Opus 5.5 often suspects it is being evaluated, which limits how well testing predicts real-world behavior. Two practical changes also appear in the announcement: thinking mode can no longer be switched off, and cybersecurity tasks are mostly re-routed to Opus 4.8.
Availability
According to the announcement, the model is available on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure, plus the Claude Platform, with zero data retention supported. Pro, Max, Team and seat-based Enterprise plans get higher limits in the five-hour usage window.
My take
What interests me most here isn’t the top of the tables, it’s the combination of price and speed. For anyone running coding agents for hours, a model that is cheaper per token and faster at output changes the math more predictably than a few benchmark points, which Anthropic itself asks readers to take with caution.
Two things to watch when migrating: always-on thinking may change latency and token usage in flows that currently turn it off, and anyone doing offensive or defensive security work needs to account for the re-routing to Opus 4.8. As always, measure on your own workload before switching your default model.