Anthropic unveiled Claude Opus 5.5 – 30% faster and 40% cheaper
For most tasks, Claude Opus 5.5 has approached the more expensive Claude Fable 5.1, while the cost of work has decreased by 40% compared to Opus 5, and generation speed has increased by more than 30%.
Prices and speed
In API, Claude Opus 5.5 costs $4 per 1 million input tokens and $20 per 1 million output tokens. For Opus 5, the rates were $5 and $25 respectively, so token prices dropped by 20%, not 40%.
| Claude Opus 5.5 | Claude Opus 5 | |
|---|---|---|
| Input tokens | $4 for 1 million | $5 for 1 million |
| Output tokens | $20 for 1 million | $25 for $1 million |
| Reading cache | $0,20 for 1 million | $0,50 for $1 million |
| Cache entry | $5 for $1 million | $6,25 for 1 million |
The claimed 40% savings apply to the cost of performing typical tasks under standard settings. Anthropic explains the difference by noting that Opus 5.5 uses fewer tokens per task, and reading from cache has become 60% cheaper. The latter is especially important for Claude Code and AI agents, which constantly use previously saved context.
At the same time, the model generates responses more than 30% faster than Opus 5. A Fast mode has also been introduced for Claude Code and the Claude Platform, capable of increasing speed by up to 2.5 times. This comes at double the cost: $8 per 1 million input tokens and $40 per 1 million output tokens.
Performance
According to internal tests by Anthropic, Opus 5.5 significantly outperforms Opus 5 in programming, computer use, and professional tasks. On Terminal-Bench 4.0, the model scored 66.4%, compared to 52.3% for Opus 5 and 55.8% for Fable 5.1. GPT-6 Astra achieved a result of 57.9%.
In GDPval-AA v2.1, which evaluates performance on real professional tasks across 44 specialties, Opus 5.5 scored 1846 Elo points. For comparison, Fable 5.1 earned 1735, Opus 5 – 1708 scored 1542, GPT-6 Astra achieved 1588, and GPT-5.6 Sol reached 1588.
| Test | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 55.8% | 52.3% | 57.9% |
| GDPval-AA v2.1 | 1846 | 1735 | 1708 | 1542 |
| AutomationBench | 40.0% | 31.4% | 26.9% | 41.4% |
| Humanity’s Last Exam with tools | 67.7% | 65.6% | 63.6% | 57.2% |
| Terminal-Bench-Science 0.1 | 58.7% | 52.6% | 29.0% | 64.6% |
| OSWorld 2.0 | 81.8% | 80.7% | 74.0% | — |
However, Opus 5.5 does not lead in all tests. GPT-6 Astra achieved a higher score on AutomationBench and Terminal-Bench-Science. Moreover, most of the reported results were provided by Anthropic itself, and at the time of the announcement, Opus 5.5 was not yet listed in the public Terminal-Bench 4.0 leaderboard. Therefore, direct comparisons should be viewed as manufacturer-provided data rather than fully independent testing.
Anthropic is separately betting on large-scale programming tasks. One of the early testers used Opus 5.5 to verify and fix a codebase of 200,000 lines in under three hours. On the same task, Opus 5 took more than 20 hours and approximately 2.5 times as many tokens.
Security and limitations
Anthropic also strengthened the model’s safeguards during autonomous code and tool usage. According to internal company estimates, Opus 5.5 attempted to breach defined environment boundaries approximately 85% less frequently than Opus 5 and Mythos 5.1. Prior to release, the model underwent additional testing by external organizations Frontier Design and METR.
Additional restrictions apply to the most sensitive tasks. In particular, a significant portion of cybersecurity-related queries may be automatically redirected to Opus 4.8. Similar measures are used for biological research and other potentially hazardous scenarios.
Opus 5.5 also uses the preserved thinking mechanism, which prevents API clients from modifying the model’s prior internal reasoning context. For new API accounts, this system takes effect on August 31, 2026. It is not possible to fully disable thinking mode in Opus 5.5.
Availability
Claude Opus 5.5 is now available in Anthropic services, Claude Code, and the Claude Platform under the identifier claude-opus-5-5, as well as via Amazon Web Services, Google Cloud, and Microsoft Azure. The company simultaneously increased five-hour usage limits for Pro, Max, Team, and corporate subscriptions with user licenses.
Opus 5.5 became the first model of the new generation. Anthropic plans to release Claude Sonnet 5.5 and Claude Haiku 5.5 in the coming weeks with similar improvements in performance, efficiency, and security.
