Anthropic ships Claude Opus 5 in Claude Code and the API

• Claude Opus 5 shipped in Claude Code and the Claude Platform with the claude-opus-5 API slug and /model claude-opus-5 in Claude Code, per ClaudeDevs' model-slug post.
• The headline coding number is 43.3% on Frontier-Bench v0.1, ahead of Fable 5 at 33.7%, according to ClaudeDevs' Frontier-Bench post.
• The migration gotcha is backward compatibility: the built-in /claude-api migrate exists, while danshipper's Day 0 test says legacy skills made Opus 5 stop early or miss instructions.
• The creative demos were browser-native and weirdly complete: LLMJunky's Rocket Arena post showed a playable 3D car game, while chrisfirst's V8 post showed a 112-part interactive engine model.
• Prompt-injection resistance was the buried security claim: bcherny's system-card note said layered defenses drove prompt-injection attack success to roughly zero.

The official launch post frames Opus 5 as near-Fable intelligence at half the price. The fun buried bits are elsewhere: Anthropic's context-engineering post says the team removed over 80% of Claude Code's system prompt for newer models, and the AWS Bedrock model card lists a 1M-token context window with 128K max output.

• Availability: Opus 5 is live in Claude Code and Claude Platform, per ClaudeDevs' launch thread.
• API and Claude Code slugs: claude-opus-5 on the API and /model claude-opus-5 in Claude Code, per ClaudeDevs' model-slug post.
• Pricing: the official launch post lists $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8.
• Product position: Opus 5 is the new default on Claude Max and the strongest model on Claude Pro, according to the official launch post.
• Effort default: Opus 5 defaults to high effort in Claude Code and Claude Platform, according to ClaudeDevs' effort post.
• Fast mode: Anthropic says fast mode runs about 2.5x faster at 2x base price, per ClaudeDevs' effort post.
• Migration tool: /claude-api migrate updates model strings and suggests Opus 5 prompt changes, according to ClaudeDevs' migration post.

First-party

• Frontier-Bench v0.1: 21.1% on Opus 4.8 to 43.3% on Opus 5, +22.2 points, per minchoi's benchmark chart.
• GDPval-AA v2: 1593 Elo on Opus 4.8 to 1861 on Opus 5, +268 Elo, per minchoi's benchmark chart.
• ARC-AGI-3: 1.5% on Opus 4.8 to 30.2% on Opus 5, +28.7 points, per minchoi's benchmark chart.
• OSWorld 2.0: 55.7% on Opus 4.8 to 70.6% on Opus 5, +14.9 points, per minchoi's benchmark chart.
• AutomationBench: 17.0% on Opus 4.8 to 26.0% on Opus 5, +9.0 points, per minchoi's benchmark chart.
• FrontierCode v1.1 Main: 46.5% on Opus 4.8 to 53.4% on Opus 5, +6.9 points, per minchoi's benchmark chart.

Third-party evaluators

• Artificial Analysis Intelligence Index: Fable 5 at 60 to Opus 5 at 61, +1 point, according to Artificial Analysis.
• GDPval-AA v2: Fable 5 at 1747 Elo to Opus 5 at 1861, +114 Elo, according to Artificial Analysis.
• AA-Briefcase: Fable 5 at 1574 Elo to Opus 5 at 1720, +146 Elo, according to Artificial Analysis.
• CursorBench 3.2 at default effort: Fable 5 at 66.5% to Opus 5 at 66.7%, +0.2 points, according to Cursor's Opus 5 availability post.

Customer-reported

• Lovable's hardest agentic coding tasks: Opus 4.7 baseline to Opus 5, +22% relative, according to a Lovable quote in Anthropic's official launch post.

• DeepSWE v1.1: Fable 5 at 69.7% to Opus 5 at 68.8%, -0.9 points, per minchoi's benchmark chart.
• DeepSWE v1.1: GPT-5.6 Sol at 72.7% to Opus 5 at 68.8%, -3.9 points, per minchoi's benchmark chart.
• FrontierCode v1.1 Main: Fable 5 at 53.5% to Opus 5 at 53.4%, -0.1 points, per minchoi's benchmark chart.
• Legal Agent Benchmark Held-out: Fable 5 at 13.3% to Opus 5 at 11.7%, -1.6 points, per minchoi's benchmark chart.
• HealthBench Professional: Mythos 5 at 66.0% to Opus 5 at 59.8%, -6.2 points, per minchoi's benchmark chart.
• Cybersecurity: Anthropic's system card says Opus 5 exceeds Opus 4.8 on cyber evals but remains below Mythos 5, especially on exploitation.
• Workflow compatibility: danshipper's Day 0 test said Opus 5 argued with existing instructions, stopped before finishing, and improved after the team deleted old skills and rebuilt from scratch.
• Effort level behavior: alliekmiller's early-access note called Opus 5 "VERY VERY eager" and said high reasoning effort had to be switched down to medium.

• Context window: the AWS Bedrock model card lists 1M tokens.
• Max output: the AWS Bedrock model card lists 128K tokens.
• Reasoning: the AWS Bedrock model card says adaptive thinking is on by default and can be disabled only with effort capped at high.
• Tooling: ClaudeDevs' tool-change post says Opus 5 supports adding or removing tools mid-conversation without invalidating the prompt cache.
• Fallback routing: ClaudeDevs' tool-change post says classifier-blocked requests route to a recommended model based on the request.
• Context engineering: trq212's context-engineering post says Anthropic removed roughly 80% of the Claude Code system prompt for its newest models.
• Prompt injection: bcherny's system-card note says Opus 5 is Anthropic's least prompt-injectable model yet, and the screenshot in that post shows Gray Swan IPI success rates of 0.2%, 1.6%, and 2.0% for k=1, k=10, and k=15 attempts.

• Old harnesses were fragile: danshipper's Day 0 test said Opus 5 behaved worse with Every's existing skills, then improved dramatically after the team started from scratch.
• Lower effort showed up twice: danshipper's Day 0 test and alliekmiller's early-access note both reported better behavior after moving down from high-effort defaults.
• Browser games were the loudest creative test: LLMJunky's Rocket Arena post showed a Rocket League-style game that consumed 27% of a 5x Max subscription, while LLMJunky's cost reply put the project cost at roughly $3.
• The same Rocket Arena project later grew multiplayer: LLMJunky's multiplayer post showed controller support and a room code for live testing.
• Opus 5 tested its own game mechanics: LLMJunky's playtest clip showed the model controlling the game to verify behavior.
• Procedural educational tools also landed: chrisfirst's V8 post described a browser V8 with 112 clickable parts, and chrisfirst's prompt asked for slider-crank geometry, cam timing, valve lift graphs, part telemetry, and no external assets.
• Motion design got an MCP-shaped demo: AmirMushich's Moda MCP demo claimed one prompt produced editable layers, keyframes, and details in five minutes.
• HTML canvas had a cleaner launch-ad test: stevibe's launch-ad demo said Opus 5 wrote the storyboard and coded a 30-second neon intro film.

• Claude surfaces: ClaudeDevs' launch thread says Opus 5 is live in Claude Code and Claude Platform, and ClaudeDevs' model-slug post gives the API and Claude Code model names.
• Linear: linear's Claude Code dropdown shows Opus 5 selectable for Coding Sessions that draft PRs from Linear.
• Cursor: Cursor's forum announcement says Opus 5 is available in the model picker and supports Zero Data Retention, unlike Fable 5.
• GitHub Copilot: GitHub's changelog says Opus 5 is available across Pro+, Max, Business, and Enterprise plans, with gradual rollout and admin enablement for business plans.
• AWS Bedrock: the Bedrock model card lists anthropic.claude-opus-5 for Bedrock runtime and Bedrock Mantle.
• Vercel AI Gateway: Vercel's changelog says claude-opus-5 is available through AI Gateway.
• Claude Code's adjacent surface is getting more agentic: minchoi's iOS simulator clip showed Claude Code building, launching, and testing an iPhone app inside the simulator.