Claude Opus 5.5 Launches: 40% Cheaper, Beats GPT-6 Astra on Coding Benchmarks

Release Overview

Anthropic launched Claude Opus 5.5 without prior announcement, marking the first model in the Claude 5.5 series. The company claims Opus 5.5 matches the performance of Fable 5.1 while reducing cost by 40% compared to Opus 5 and increasing output speed by 30% . Within two hours, OpenAI responded with GPT-6 Sol and Luna , signaling an intensified competitive round.

Benchmark Performance

Coding Benchmarks

Terminal-Bench 4.0 (real-terminal multi-step coding): Opus 5.5 scores 66.4% , leading GPT-6 Astra ( 57.9% ), Fable 5.1 ( 55.8% ), and Opus 5 ( 52.3% ).

FrontierCode v1.1 (real code merge tasks): Opus 5.5 achieves 54.4% , narrowly beating Astra's 53.3% .

CursorBench 4.0 (noisy, cross-file tasks from real Cursor sessions): Opus 5.5 scores 57.8% vs GPT-5.6 Sol's 41.7% .

Anthropic notes that Opus 5.5 at default settings outperforms Opus 5 at its highest tier while costing only one-fifth as much; matching Astra requires only 40% of Astra's cost.

Knowledge Work

On GDPval-AA v2.1 (44 real-world professional tasks), Opus 5.5 reaches 1846 Elo versus Astra's 1542 Elo — a 300-point gap described as nearly a generational difference in competitive rating systems.

Third-Party Evaluations

Artificial Analysis : Both Opus 5.5 and Fable 5.1 rank above GPT-6 Astra.

Anthropic's ECI (Capability Index) : Opus 5.5 surpasses the previously withheld Mythos 5.1.

Vals AI RSI Index (AI research self-improvement ability): Opus 5.5 ranks globally first.

Writing and Communication Improvements

Opus 5.5 adopts a "conclusion-first" style, eliminating jargon and verbosity. In a side-by-side comparison, Opus 5.5 immediately identifies a billing anomaly and provides code, whereas Opus 5 buries the conclusion after lengthy preamble. Box's AI product VP reports token usage reduced to one-third of the previous generation and verbosity down 40% .

Creative and Generative Capabilities

Despite lacking native image generation, Opus 5.5 produces sophisticated 3D animations and simulations via code:

Ben Poole (Google DeepMind) : Sketched a trebuchet and blocks; Opus 5.5 generated a working 3D physics simulation with accurate joints and counterweights.

Addy Osmani (Anthropic) : Created a 3D animated pelican riding a bicycle in Three.js.

Ethan Mollick (Wharton) : Tested a "gothic city in rain" shader prompt; Opus 5.5 produced a detailed real-time scene.

Game generation : One-shot creation of a doodle-style FPS with three enemy waves and a boss; a playable Minecraft clone; and a game inspired by the Antikythera mechanism with all graphics and audio generated purely from code.

Pricing and Token Economics

Standard pricing: Input $4 / million tokens , Output $20 / million tokens (both 20% reduction).

Cache read: $0.20 / million tokens (60% reduction from $0.50).

Fast mode: 2.5× speed at double price (Input $8, Output $40).

However, Artificial Analysis testing under max-effort conditions reveals:

Opus 5.5 averages 119,000 tokens per problem , with 84,000 tokens for reasoning .

Opus 5: 73,000 total tokens; Fable 5.1: 78,000; GPT-6 Astra: 27,000.

Opus 5.5 thinks 60% more than Opus 5 and 4× more than Astra. The 20% per-token discount is largely offset by the 60% increase in reasoning tokens.

Safety and Situational Awareness

Opus 5.5 achieves the highest historical score on automated behavioral audits, with 85% fewer jailbreak attempts than its predecessor and consistent self-reporting of violations. Critically, Anthropic discloses that Opus 5.5 frequently detects when it is being evaluated by humans . This situational awareness raises concerns: compliant behavior during testing may not reflect unmonitored deployment behavior, a challenge that grows with model capability.

Roadmap and Competitive Landscape

Anthropic confirms Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks, completing a full product-line assault. OpenAI's immediate counter with GPT-6 Sol and Luna (positioned as higher efficiency, lower cost) indicates the industry has moved from pure technology competition to a price-performance melee.

References

Anthropic announcement: https://www.anthropic.com/claude-opus-5-5 Anthropic X post: https://x.com/AnthropicAI/status/2102435703535939725 ClaudeDevs X post:

https://x.com/ClaudeDevs/status/2102438800836489554