Opus 4.8 ships at Opus 4.7 pricing, with the first 0% score on uncritically reporting flawed results in the Claude line and a USAMO math benchmark that jumped 27 points overnight.
Anthropic released Claude Opus 4.8 today. The model's honesty benchmarks are the story: it scores 0% on uncritically accepting and reporting flawed results, the first Claude model to reach that mark, and shows a more than tenfold reduction in overconfidence compared to 4.7. The USAMO 2026 math benchmark moved from 69.3% to 96.7%. SWE-bench Pro reached 69.2% from 64.3%. Source: anthropic.com/news/claude-opus-4-8, also reported by The Decoder and TechCrunch.
Pricing holds: $5/$25 per MTok. Context window: 1M tokens on Claude API, Amazon Bedrock, and Google Cloud Vertex AI. 200k on Microsoft Foundry. Max output: 128k tokens. Fast mode drops to 2x standard pricing for 2.5x speed, down from 6x previously.
Anthropic's new flagship. Available on Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Context window 1M tokens (200k on Foundry). Max output 128k tokens. Benchmarks vs 4.7: SWE-bench Pro 69.2% (from 64.3%), SWE-bench Verified 88.6% (from 87.6%), USAMO 2026 96.7% (from 69.3%), GraphWalks long-context F1 at 1M tokens 68.1% (from 40.3%). First Claude model to score 0% on uncritically reporting flawed results. Overconfidence more than tenfold lower than 4.7. Pricing unchanged: $5/$25 per MTok. Fast mode: 2x standard rate for 2.5x speed.
client.messages.create(
model="claude-opus-4-8",
max_tokens=1024,
messages=[{"role": "user", "content": "..."}]
)The Claude API now accepts role: "system" messages after a user turn, on Opus 4.8. Previously, system prompts were fixed at conversation start. Enables dynamic system prompt updates mid-session without restarting the conversation. stop_details field now documented publicly.
Pass a message with role: "system" at any position in your messages array after the first user turn. Requires claude-opus-4-8.
Opus 4.8 defaults to high effort in Claude Code. Dynamic workflows (research preview): describe a large task and Claude will spin up tens to hundreds of background subagents to execute it in parallel. Fast mode for Opus 4.8 priced at 2x standard rate (2.5x speed). Lean system prompt is now the default for all models except Haiku, Sonnet, and Opus 4.7 and earlier. /simplify now runs cleanup-only review and applies fixes automatically. Shell background sessions: type ! <command> to run a shell command as a background session with attach/detach capability. Streaming tool execution always enabled including on Bedrock, Vertex, and Foundry. Worktree isolation guard prevents subagents from bypassing restrictions.
claude update to get v2.1.154. Dynamic workflows: start a large task description and Claude will propose splitting across agents if applicable.
Claude.ai users can now select the effort level Claude applies to a response. Ships alongside Opus 4.8. High effort is the default for Opus 4.8. The control surfaces the tradeoff between quality and response speed per conversation.
Reported by GitHub. Opus 4.8 is now selectable for all GitHub Copilot users in Copilot Chat, inline suggestions, and agent mode. Previously available in limited preview.
v0.105.0: adds claude-opus-4-8 model support, mid-conversation system blocks, and usage.output_tokens_details. v0.105.1: switches PyPI releases to Trusted Publishing for improved supply-chain security. v0.105.2: minor patch. Three releases inside 24 hours.
pip install --upgrade anthropicv0.100.0: adds claude-opus-4-8, mid-conversation system blocks, and usage.output_tokens_details. v0.100.1: fixes streaming to carry encrypted_content on beta compaction blocks. Dependency lockfiles updated across foundry-sdk and aws-sdk sub-packages.
npm install @anthropic-ai/sdk@latestNo research paper published. No system prompt update detected (docs.anthropic.com returned 403 throughout sweep). No deprecation notice. No standalone Claude Apps update beyond the effort-control feature attached to the Opus 4.8 launch.
Daily digest at 9 PM ET. Weekly magazine every Friday morning. Six labs, one feed. No spam, one-click unsubscribe.