Claude Opus 4.8
Anthropic's top general-availability workhorse for complex agentic coding and enterprise work.
Context window
1M
Input / 1M tokens
$5.00
Output / 1M tokens
$25.00
Provider
Anthropic
Standard pricing $5/$25 per 1M input/output tokens with the full 1M context included at no surcharge. Fast mode (research preview) costs $10/$50 per 1M for ~2.5x higher output tokens/sec. Up to 90% savings with prompt caching (minimum cacheable prompt lowered to 1,024 tokens) and 50% with batch processing. · Data verified 2026-07-24
Claude Opus 4.8 (May 2026) builds on Opus 4.7 with stronger long-horizon agentic coding, better effort calibration, and improved tool triggering. It adds a research-preview fast mode that is roughly 3x cheaper than prior fast modes, effort control across surfaces, mid-conversation system messages, and notably higher honesty (about 4x less likely to let code flaws pass unremarked than Opus 4.7). It uses adaptive thinking only - manual thinking budgets and non-default sampling parameters return 400 errors.
Capability index
Relative estimates (0-100) to place this model against its peers, grounded in published benchmarks.
How to access it
Available via the Claude API, Amazon Bedrock, Google Cloud (Vertex AI), and Microsoft Foundry, plus claude.ai and Cowork. Anthropic's recommended default for complex agentic coding and enterprise work.
Strengths
- ✓Best-in-GA agentic coding: 88.6% SWE-bench Verified and 69.2% SWE-bench Pro
- ✓Improved long-context handling with fewer compactions and better compaction recovery
- ✓Recalibrated effort levels (medium/high/xhigh) for precise cost-vs-thoroughness control
- ✓Higher honesty - far less likely to let code flaws pass unflagged
- ✓Fast mode option for ~2.5x throughput and a lower 1,024-token cache minimum
Best for developers who...
When to choose it (and when not to)
Reach for Claude Opus 4.8 when...
- →You need top-tier autonomous coding and multi-step agentic reliability
- →You run long-horizon agent traces that previously derailed after compaction
- →You want enterprise-grade reasoning with 1M context and 128k output
- →You need throughput and are willing to pay fast-mode premium for 2.5x speed
Look elsewhere if...
- ✕You are cost-sensitive at scale - Sonnet 5 delivers close performance for far less
- ✕You require manual extended-thinking budgets or custom sampling parameters (400 error)
- ✕You want the single strongest available model regardless of price (consider Fable 5)
How to use it
- ›Use the effort parameter (default 'high') with adaptive thinking; re-baseline cost/latency since effort levels were recalibrated vs Opus 4.7
- ›Leverage mid-conversation system messages to update instructions without breaking prompt-cache hits on earlier turns
- ›Remove temperature/top_p/top_k and manual thinking budgets - both return 400
- ›Take advantage of the 1,024-token cache minimum to cache shorter shared prefixes
Quickstart
Pythonimport anthropic
client = anthropic.Anthropic()
response = client.messages.create(
model="claude-opus-4-8",
max_tokens=16000,
thinking={"type": "adaptive"},
output_config={"effort": "high"},
messages=[
{"role": "user", "content": "Diagnose and fix the failing integration test."}
],
)
print(response.content[0].text)Manual thinking budgets and non-default temperature/top_p/top_k return a 400 error; control depth via effort (low/medium/high/xhigh).
API model id: claude-opus-4-8
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| SWE-bench Verified | 88.6% | vs Opus 4.7 87.6% |
| SWE-bench Pro | 69.2% | Leads active models; ~5 pts over Opus 4.7 |
| Terminal-Bench 2.1 | 74.6% | |
| GPQA Diamond | 93.6% | |
| Artificial Analysis Intelligence Index | 61.4 | Edges GPT-5.5 at 60.2 |
Tools powered by Claude Opus 4.8
Compare Claude Opus 4.8
Claude Opus 4.8 vs Gemini 3.5
Google DeepMind - 1,048,576 tokens (Gemini 3.5 Flash; Pro variant not yet released) ctx
Claude Opus 4.8 vs GPT-5.6 Sol
OpenAI - 1500000 ctx
Claude Opus 4.8 vs Claude Sonnet 5
Anthropic - 1M ctx
Compare Claude Opus 4.8 with any other model
Build a comparison →All model comparisons →