Claude Opus 4.8
Released: May 28, 2026
Model ID: claude-opus-4-8
Pricing: $5 / $25 per million input / output tokens (unchanged from 4.7)
Status: legacy. Superseded by claude-opus-5-5.md. Still the required layer
when upgrading from Opus 4.7 or older.
Last verified: legacy, written at the Opus 4.8 launch (May 2026)
Changelog (vs Opus 4.7)
- Effort control: User-selectable effort levels in Claude.ai, Cowork, and
API. Controls how much compute Claude applies per task. Token allocation
recalibrated — re-baseline any tuned effort settings.
- Dynamic workflows (research preview): Claude Code can plan complex tasks
and spawn hundreds of parallel subagents in a single session. Useful for
large-scale refactors, migrations, and repo-wide analysis.
- Mid-task instructions: Messages API accepts
role: "system" entries
within the messages array (after a user turn). Update instructions mid-task
without breaking prompt cache or routing through a user turn.
- Lower prompt cache minimum: Reduced to 1,024 tokens (down from prior
threshold).
- Fast mode: 2.5x speed, now 3x cheaper than previous versions. Available
as research preview on the API.
- Refusal details:
stop_details field on refusal responses is now public.
Returns category (cyber, bio, or null) plus explanation.
- Sharper judgment and reduced hallucinations in long-running agentic tasks.
- Gains on SWE-Bench Pro, Online-Mind2Web, and multidisciplinary reasoning.
- Better honesty — less likely to make unsupported claims.
- More effective at managing autonomous, long-running workflows.
Migration from Opus 4.7
Drop-in upgrade. No breaking API changes.
| Step |
Action |
| Model ID |
Change to claude-opus-4-8 |
| Effort settings |
Re-baseline — token allocation recalibrated |
| Mid-task instructions |
Optional — new capability, no action needed |
| Fast mode |
Optional — available if latency is a priority |
Migration from Opus 4.6 or earlier
Must apply all Opus 4.7 breaking changes first, then update model ID.
Opus 4.7 breaking changes (required)
| Change |
Detail |
| Sampling parameters |
Setting temperature, top_p, or top_k to non-default values returns 400. Remove them. |
| Extended thinking |
Replace manual extended thinking config with {type: "adaptive"} and effort parameter. |
| Tokenizer |
New tokenizer — re-evaluate cost estimates and token limits. |
| Assistant prefills |
Assistant message prefills return 400. Use structured outputs or output_config.format instead. |
After applying these, proceed with the 4.7 → 4.8 migration above.
Quirks and best practices
- Literal instruction following: Opus 4.8 (inherited from 4.7) follows
instructions more literally. Vague or implicit instructions may produce
different results than on 4.6. Be explicit about format, verbosity, and
scope.
- Effort recalibration: If you tuned effort levels on 4.7, retest on 4.8.
The token budget per effort level has changed.
- Mid-task instruction pattern: Use mid-conversation system messages to
inject updated context (e.g., new files discovered, scope changes) without
restarting the prompt. Preserves cache hits.
- Dynamic workflow orchestration: For large migrations, leverage dynamic
workflows to parallelize discovery, classification, and implementation across
subagents. Each subagent gets its own context window.
- Refusal transparency: Monitor
stop_details on safety refusals to
distinguish content-policy blocks from model uncertainty.
- Prompt caching: With the lower 1,024-token minimum, cache even smaller
system prompts for cost savings on high-volume workloads.
Model ID reference
| Version |
Model ID |
Status |
| Opus 4.8 |
claude-opus-4-8 |
Current (GA) |
| Opus 4.7 |
claude-opus-4-7 |
Active |
| Opus 4.6 |
claude-opus-4-6 |
Active |
| Opus 4.5 |
claude-opus-4-5 |
Active |
| Opus 4.1 |
claude-opus-4-1 |
Active |
| Opus 4 |
claude-opus-4-20250514 |
Deprecated soon |