On September 22, 2026, Anthropic officially announced Claude Opus 5.5, the first model in its new generation and its latest flagship AI, making it available immediately. This marks the first major model release since the company introduced its 2026 principle, "We Must Pace the Frontier." While maintaining advanced reasoning capabilities on par with its higher-tier model, Claude Fable 5.1, Opus 5.5 cuts operational costs for standard workloads by 40% compared to its predecessor, Claude Opus 5, and boosts generation speed by more than 30%. Claude Sonnet 5.5 and Claude Haiku 5.5 are expected to follow within weeks.
The Structure Behind the 40% Cost Reduction: Sweeping Price Cuts and Expanded Cache Discounts
The biggest change in Opus 5.5 comes from a comprehensive overhaul of API pricing and a significant expansion of prompt caching discounts. The API list price is now set at $4 per million input tokens and $20 per million output tokens—a 20% reduction on both counts from Opus 5's pricing ($5 input / $25 output).
Furthermore, the price for cached prompt reads—which account for the bulk of costs in agentic operations and iterative coding tasks—has been revised down to $0.20 per million tokens. That's a 60% cut from Opus 5's $0.50, representing a 95% discount relative to the non-cached input rate of $4. The price for cache writes with a 5-minute time-to-live (TTL) has also been lowered to $5 per million tokens, down from Opus 5's $6.25.
API Pricing Comparison: Claude Opus 5.5 vs. Opus 5
| Pricing Category (per million tokens) | Claude Opus 5.5 | Claude Opus 5 | Change | Effect |
|---|---|---|---|---|
| Input Price | $4 | $5 | 20% reduction | Lower base input cost |
| Output Price | $20 | $25 | 20% reduction | Lower generated-token cost |
| Cache Read Price | $0.20 | $0.50 | 60% reduction | 95% discount vs. input rate |
| Cache Write Price | $5 | $6.25 | 20% reduction | Cheaper context initialization |
- Claude Opus 5.5
- Claude Opus 5
データを表で見る
| Claude Opus 5.5 (USD) | Claude Opus 5 (USD) | |
|---|---|---|
| Input Price | 4 | 5 |
| Output Price | 20 | 25 |
| Cache Write | 5 | 6.25 |
| Cache Read | 0.2 | 0.5 |
Combined, the lower per-token rates and improved token efficiency for completing tasks add up to an effective 40% cost reduction in standard real-world operations.
Alongside this, a new developer-facing high-speed option called "Fast mode" has launched on Claude Code and the Claude Platform. Fast mode delivers up to 2.5x faster output, priced at $8 per million input tokens and $40 per million output tokens. It retains the 1-million-token context window (supporting both text and image inputs) and supports Zero Data Retention. Across the Pro, Max, and Team subscription plans, 5-hour usage limits have been relaxed, and a new feature lets users bank rate-limit resets to trigger at a time of their choosing.
New Benchmark Leadership in Agentic Coding and Knowledge Work
In terms of reasoning performance, Opus 5.5 has demonstrated standout results in autonomous terminal operation and large-scale codebase modification tasks. On Terminal-Bench 4.0, which measures autonomous problem-solving in command-line environments, Opus 5.5 (xhigh effort) scored 66.4%—clearly ahead of OpenAI's GPT-6 Astra (high effort: 57.9%), Claude Fable 5.1 (55.8%), and its predecessor Opus 5 (52.3%).
On FrontierCode v1.1 (Main), a software development benchmark involving long-context handling, Opus 5.5 achieved 54.4%, beating GPT-6 Astra (53.3%) and Fable 5.1 (50.3%). On CursorBench 4.0, it scored 57.8%, opening a wide gap over GPT-5.6 Sol (41.7%) and Opus 5 (46.6%).
Benchmark Comparison Among Leading Frontier Models
| Benchmark | Evaluation Domain | Claude Opus 5.5 | Claude Fable 5.1 | GPT-6 Astra | Claude Opus 5 |
|---|---|---|---|---|---|
| Terminal-Bench 4.0 | Autonomous terminal operation (official max) | 66.4% | 55.8% | 57.9% | 52.3% |
| FrontierCode v1.1 | Agentic development (Main) | 54.4% | 50.3% | 53.3% | - |
| CursorBench 4.0 | Long-session editor reasoning | 57.8% | 51.8% | - | 46.6% |
| GDPval-AA v2.1 | Real-world work intelligence (Elo score) | 1846 | 1735 | 1542 | 1708 |
- Claude Opus 5.5
- Claude Fable 5.1
- GPT-6 Astra
- Claude Opus 5
データを表で見る
| Claude Opus 5.5 (%) | Claude Fable 5.1 (%) | GPT-6 Astra (%) | Claude Opus 5 (%) | |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 66.4 | 55.8 | 57.9 | 52.3 |
| FrontierCode v1.1 | 54.4 | 50.3 | 53.3 | — |
| CursorBench 4.0 | 57.8 | 51.8 | — | 46.6 |
One real-world-scale case study involved a complete migration of a web server from C to Rust. Both Opus 5.5 and Fable 5.1 produced code that passed the regression test suite, but Opus 5.5 showed an edge in both time-to-completion and generation cost. Early testers also reported cases where engineering teams completed large-scale code migration tasks—which would normally take a long time—in remarkably short order.
On GDPval-AA v2.1, which measures real-world knowledge work, Opus 5.5 scored 1846 Elo, well ahead of Fable 5.1 (1735), GPT-6 Astra (1542), and GPT-5.6 Sol (1588). On AutomationBench, a business-automation benchmark run by Zapier, it scored 40.0% (versus 41.4% for GPT-6 Astra). Since AutomationBench was run without a fallback model, triggered safeguards were counted as incorrect answers—meaning real-world scores are likely to be even higher.
Intelligence Index Leadership and What the Pareto Frontier Reveals About Real Costs
An extensive external audit by the independent evaluation firm Artificial Analysis confirmed Opus 5.5's objective standing in the market. On the firm's composite Artificial Analysis Intelligence Index, Opus 5.5 (max effort) posted a score of 58—the highest score recorded by any frontier model measured to date, putting it in sole first place.
Of the ten individual evaluations that make up the Intelligence Index, Opus 5.5 took the top spot in six. It scored 61.4% on the extremely difficult Humanity's Last Exam benchmark and 66.9% on SciCode, a scientific code-generation benchmark. In Artificial Analysis's own independent measurement of Terminal-Bench 4.0, it reached 59.6%, tying with the leader GPT-6 Astra (xhigh) and improving 11 points over Opus 5.
Particularly notable is AA-Briefcase v1.1, a private, high-difficulty real-work evaluation built on the open-source agent execution framework Stirrup. Opus 5.5 scored 1822 Elo here, a gain of 143 points over Fable 5.1. Beyond depth of analysis, it's also the first Anthropic model to surpass GPT-5.6 Sol in presentation quality of its outputs.
Balancing Increased Output Tokens Against Per-Task Cost
Data from Artificial Analysis reveals an important fact about the economics of reasoning models. Opus 5.5 (max) generates an average of roughly 119,000 output tokens per task—far more than Opus 5 (max: about 73,000 tokens), Fable 5.1 (max: about 78,000 tokens), or GPT-6 Astra (max: about 27,000 tokens), reflecting deeper deliberation.
Normally, an increase in token usage would translate directly into higher costs. But for Opus 5.5, the 20% cut in per-token pricing and the 60% cut in cache-read pricing take effect, keeping the total cost per task on par with Opus 5.
Moreover, across the five reasoning-effort settings (low, medium, high, xhigh, max), the four settings from medium through max all sit on the Pareto frontier of "intelligence versus cost per task." Compared with other leading frontier models, Opus 5.5 maintains a position where it either delivers equivalent performance at lower cost, or higher intelligence at equivalent cost.
Moving Beyond "Claudish" Writing: A Refresh in Conversational Quality
Alongside its capability gains, Opus 5.5 brings a major change that users will notice directly: improvements to the model's writing tone and conversational structure. Previously, Opus 5 drew criticism from the user community for being overly ornate, roundabout, and prone to distinctive quirks—a phenomenon some dubbed "Claudish writing."
Opus 5.5 addresses this tendency head-on. It now leads with conclusions and key information up front, organizing structure logically for a more direct, clear style of writing. Early testers described the change as making the model's output "easier to follow, and more natural—like something I'd write myself."
Beyond improving readability, this shift also has practical safety implications. When autonomous agents report complex work plans or the reasoning behind code changes to developers, a structure that makes key points immediately clear reduces the burden on human reviewers and helps prevent oversights—acting as a safeguard in its own right.
Layered Safeguards Supporting Robust Autonomous Agent Operation
With an eye toward long-duration autonomous agent operation in enterprise settings, Anthropic has significantly strengthened its safety management and attack-resistance mechanisms. Pre-release testing included third-party evaluations by Frontier Design and METR (Model Evaluation and Threat Research).
In an automated behavioral audit run across roughly 2,000 simulated environments, Opus 5.5 achieved the best alignment score of any recent Claude model. Instances of "containment evasion"—where an agent recognizes and attempts to circumvent the limits of a monitoring sandbox—dropped by about 85% compared to both Opus 5 and Claude Mythos 5.1. All minor attempts that were identified fell into the low-severity category and were self-reported by the model. In a prompt injection resistance evaluation conducted by AI security firm Gray Swan, Opus 5.5 tied with Fable 5.1 for the lowest attack success rate among all models tested.
Transparent Fallback for Dual-Use Domains
For advanced specialized domains where misuse risk is a concern, Anthropic has built in an automatic fallback mechanism for production environments. Cybersecurity-related tasks are automatically routed to Claude Opus 4.8, while tasks involving biology or frontier LLM development that carry bioweapon risk are routed to Claude Opus 5.
To avoid disrupting the legitimate work of research institutions and defense-sector professionals, organizations that pass a rigorous vetting process are granted expanded access with relaxed restrictions through the Life Sciences Verification Program and the Cyber Verification Program (which offers three tiers of permissions).
Defenses Against Thought Tampering and Distillation Attacks
Anthropic has also shored up defenses against "distillation attacks"—where an attacker systematically extracts a model's reasoning process via API using tens of thousands of accounts to harvest training data for a competing model. The "preserved thinking" feature first introduced in Fable 5.1 is now fully applied in Opus 5.5, blocking users from deliberately tampering with the reasoning process embedded in prior conversation context, thereby preserving the integrity of the model's reasoning.
Additionally, in Opus 5.5, thinking mode can no longer be fully disabled—output must pass through the reasoning process. EU AI Act-compliant digital watermarking has also been applied as a standard feature.
As frontier AI increasingly takes on autonomous software development and research work, the criteria for choosing a model have shifted away from simple benchmark score competitions toward a broader combination of factors: token economics, the naturalness of collaboration—including writing style—and the overall discipline of operating without deviation. The balance of intelligence and cost structure that Claude Opus 5.5 has presented is poised to become a new benchmark in the primary battleground of agentic operations.