Anthropic has introduced Claude Opus 5.5 — the first model in the Claude 5.5 family. According to the company, it outperforms Opus 5 in agentic programming, computer use, and knowledge work while running 30%+ faster and roughly 40% cheaper on typical tasks. The price of access to the flagship tier via the API has dropped to $4 per 1M input tokens and $20 per 1M output tokens, compared to $5/$25 for Opus 5, and cache reads have become cheaper, dropping from $0.50 to $0.20 per 1M. Anthropic has also raised the 5-hour usage limits on Pro, Max, Team, and Enterprise plans and confirmed that Claude Sonnet 5.5 and Haiku 5.5 will be released in the coming weeks.
What happened
Anthropic has released Claude Opus 5.5 — the first model in the Claude 5.5 line, already available via the API and in Claude Code, meaning this is a deployable change, not a research demo announcement. The company claims superiority over Opus 5 in agentic programming, computer use, and knowledge work, while simultaneously accelerating by more than 30% and reducing costs by roughly 40% on typical tasks. API pricing: $4 per 1M input tokens, $20 per 1M output tokens, and $0.20 per 1M for cache reads; for Opus 5, these same items cost $5, $25, and $0.50. For latency-sensitive scenarios, an accelerated Fast mode has been introduced — in Claude Code and on the Anthropic Platform, it runs up to 2.5x faster at $8 and $40 per 1M input and output tokens, respectively. In vendor benchmarks, the model scored 66.4% on Terminal-Bench 4.0, 54.4% on FrontierCode v1.1, and 67.7% on Humanity's Last Exam with tools; according to Anthropic, it outperforms Opus 5 and GPT-6 Astra on most agentic tests. At the same time, the company has raised the 5-hour usage limits on Pro, Max, Team, and Enterprise plans and confirmed that Claude Sonnet 5.5 and Haiku 5.5 will appear in the coming weeks.
Context
To gauge the scale, the dynamics of the previous generation are important: Opus 5 cost $5/$25 per 1M tokens and $0.50 for cache reads, and its price was considered a benchmark for the flagship class. Opus 5.5 lowers this bar for the first time — and, it seems, intentionally: the typical Anthropic pattern is that the flagship paves the price trail, and the junior models in the family consolidate it, so the release of Sonnet 5.5 and Haiku 5.5 will likely extend the new price-to-performance ratio across the entire price stack. The company's own rhetoric is also telling: Anthropic openly warns that at such capability levels, the gap with Claude Fable 5.1 in real tasks is smaller than benchmarks suggest. This is a rare admission in the industry that vendor numbers on synthetic tests are a weak proxy for real work, and simultaneously confirms that benchmark inflation has become a systemic problem that developers themselves no longer hide.
Why this matters for the industry
The main change for the industry is economics: the flagship capability tier has become cheaper, dropping from $5/$25 to $4/$20 per 1M tokens, and cache reads from $0.50 to $0.20, meaning the cost of agentic pipelines and long-running tasks is directly and verifiably reduced, without relying on marketing phrasing. For teams building products on Claude, this expands the set of unit-economically viable scenarios: what previously didn't add up in cost-per-task may pass after recalculation. Fast mode adds a "pay for speed" option for interactive UX scenarios where wall-clock time is more critical than price. At the same time, the price drop at the flagship tier creates pressure on competitors, who now have to justify their own price lists, and the announcement of Sonnet 5.5 and Haiku 5.5 means that the main production load in the coming weeks will likely shift to cheaper models in the family, while Opus 5.5 will remain for complex agentic cycles. If the trend of "each flagship generation is significantly cheaper and faster" continues, agentic workflows risk becoming a standard product layer, and value will shift from model access to orchestration and result verification.
Why this matters for users
For those working with Claude Code, the API, or using Claude for large codebases and long tasks, the changes are already tangible: existing pipelines can be migrated from Opus 5 to Opus 5.5 without code changes, and after their own regression checks, they can achieve roughly 40% savings and 30%+ acceleration on typical tasks. The raised 5-hour limits on Pro, Max, Team, and Enterprise reduce friction for teams working within subscriptions, and Fast mode offers a choice for those who value response speed. Anthropic also claims the model is more resilient to prompt injection. The company provides illustrative examples: migrating 680,000 lines of code in less than a day and rewriting HAProxy from C to Rust in 9.5 hours, passing almost all of the project's regression tests. Model responses, according to the vendor, have become shorter, reducing both token consumption and wait times.
What is still unknown / limitations
Most capability claims currently rely solely on vendor benchmarks: absolute values like 66.4% on Terminal-Bench 4.0, 54.4% on FrontierCode v1.1, and 67.7% on Humanity's Last Exam are given without Opus 5 baselines, eval configurations, and confidence intervals, and the FrontierCode result, without knowing the distribution of other models, is poorly informative, as is the choice of benchmarks without full tables. The headline cases — the 680,000-line code migration and the HAProxy port to Rust — are single announced examples without a reproduction protocol: n=1, selection criteria are unclear, and they prove the quality of individual runs rather than a systemic leap. Anthropic itself admits that the gap with Claude Fable 5.1 in real tasks is smaller than benchmarks suggest — honest, but this is indirect confirmation of inflated numbers. Independent evals have not yet accumulated, and only after they appear will it become clear how well vendor metrics align with practice; for self-hosted and on-prem scenarios, the release changes nothing.
Sources
Author
Look at AI, editorial team