Claude Opus 4.8 launched on May 28, 2026, six weeks after Opus 4.7. Same price, same context window, same API structure — but measurably better at coding, more honest about its limitations, and paired with a new dynamic workflows feature that changes what Claude Code can tackle. Here is everything that changed and whether you should switch immediately.
The short answer: yes, upgrade now. There is no downside.
Benchmark improvements
| Benchmark | Opus 4.7 | Opus 4.8 | Improvement |
|---|---|---|---|
| SWE-bench Pro | 64.3% | 69.2% | +4.9 points |
| Terminal-Bench 2.1 | 65.8% | 74.2% | +8.4 points |
| SWE-bench Verified | 85.2% | 88.6% | +3.4 points |
| OSWorld-Verified | 82.3% | 87.1% | +4.8 points |
| Humanity’s Last Exam | 51.2% | 57.9% | +6.7 points |
| Finance Agent v2 | 48.7% | 53.9% | +5.2 points |
| Artificial Analysis Index | 57.3 | 61.4 | +4.1 points |
Every benchmark improved. The largest gain is Terminal-Bench (+8.4 points), which measures command-line task completion — directly relevant to Claude Code users. SWE-bench Pro’s +4.9 points means Opus 4.8 resolves nearly 5% more real GitHub issues than its predecessor.
Honesty: the headline improvement
Anthropic’s internal evaluations show Opus 4.8 is four times less likely to let flawed code pass without flagging the issue. In practice, this means:
- It catches its own mistakes before reporting results
- It pushes back when a plan is not sound
- It flags uncertainties rather than making unsupported claims
- It proactively identifies issues with inputs and outputs
This was the most common complaint about Opus 4.7 — it would confidently claim progress despite thin evidence. Opus 4.8 addresses this directly. Harvey’s Head of Applied Research noted it’s “the first model to break 10% overall on the all-pass standard” for their Legal Agent Benchmark.
New features in 4.8
Dynamic workflows (Claude Code)
The biggest new capability. Claude can now plan large tasks, spawn hundreds of parallel subagents, and verify results before reporting back. This enables:
- Codebase-wide migrations across hundreds of thousands of lines
- Security audits that search a service in parallel
- Language ports (Bun was ported from Zig to Rust in 11 days using this)
- Bug hunts with independent verification on every finding
Available on Max, Team, and Enterprise plans. See our dynamic workflows guide for setup instructions.
Effort control (all products)
A new control alongside the model selector lets you choose how much effort Claude puts into a response:
| Level | Use case | Token usage |
|---|---|---|
| Low | Quick answers, simple tasks | Minimal |
| Medium | Balanced | Moderate |
| High (default) | Complex tasks, similar tokens to 4.7 | Standard |
| Extra/xhigh | Difficult problems, long workflows | Higher |
| Max | Hardest problems | Maximum |
In Claude Code, use /effort to switch. In the API, configure via the thinking parameter.
System messages mid-conversation
The Messages API now accepts system entries inside the messages array. You can update Claude’s instructions mid-task without breaking the prompt cache. Use cases:
- Update permissions as an agent runs
- Adjust token budgets mid-session
- Change environment context dynamically
This was previously impossible without starting a new conversation or wasting a user turn.
Fast mode: 3x cheaper
Fast mode (2.5× speed) pricing dropped dramatically:
| Opus 4.7 Fast | Opus 4.8 Fast | Change | |
|---|---|---|---|
| Input | $30/M | $10/M | 3× cheaper |
| Output | $150/M | $50/M | 3× cheaper |
This makes fast mode practical for production use. Previously it was prohibitively expensive for most workloads.
What stayed the same
- Pricing: $5 input / $25 output per million tokens (unchanged)
- Context window: 1,000,000 tokens
- API compatibility: Drop-in replacement, just change model string to
claude-opus-4-8 - Availability: Same platforms (API, Bedrock, Vertex AI, Foundry)
- Vision: Still supports high-resolution image input
Tool calling improvements
Cursor’s CEO reported that tool calling is “meaningfully more efficient, using fewer steps for the same intelligence.” Devin’s CEO confirmed it “fixes the comment-verbosity and tool-calling issues we saw with Opus 4.7.”
In practice, this means:
- Fewer API calls to complete the same task
- Lower total token usage in agentic workflows
- Better coherence across long tool-calling chains
Migration guide
Switching from 4.7 to 4.8 requires one change:
# Before
response = client.messages.create(model="claude-opus-4-7", ...)
# After
response = client.messages.create(model="claude-opus-4-8", ...)In Claude Code, the model updates automatically. In the API, change the model string. No other code changes needed.
If you are using the claude-opus-latest alias, it now points to 4.8 automatically.
Should you upgrade?
Yes, immediately. There is no tradeoff. Same price, better at everything, fewer errors, new features. The only reason to stay on 4.7 is if you have a production system that has been extensively tested against 4.7’s specific behavior and you need time to re-validate.
For most users, switching is a one-line change with immediate benefits.
What’s coming next
Anthropic confirmed two things:
- Lower-cost models with “many of the same capabilities as Opus” are in development
- Mythos-class models with higher intelligence than Opus are coming “in the coming weeks”
Mythos Preview is currently limited to cybersecurity work. Broader release depends on safety guardrails being finalized.
FAQ
Is Opus 4.8 a major upgrade or a minor one?
Anthropic calls it “modest but tangible.” The benchmarks show 4-8 point improvements across the board. The honesty improvement (4× fewer unflagged errors) is the most impactful change for daily use. Dynamic workflows are a major new capability but only available in Claude Code.
Do I need to change my prompts?
No. Opus 4.8 is a drop-in replacement. Your existing prompts, system messages, and tool definitions work unchanged. You may notice better results without any prompt changes.
Will Opus 4.7 be deprecated?
Anthropic has not announced a deprecation date. Both models are available simultaneously. However, there is no reason to use 4.7 over 4.8.
Does the effort control affect pricing?
No. You pay the same per-token rate regardless of effort level. Higher effort means more tokens used (and thus higher total cost), but the rate per token is unchanged.
How does Opus 4.8 compare to GPT-5.5?
Opus 4.8 leads on SWE-bench Pro (69.2% vs 58.6%) and the overall intelligence index. GPT-5.5 has a higher Terminal-Bench score with its native harness. For a detailed comparison, see Opus 4.8 vs GPT-5.5.
Is Opus 4.8 worth the price vs cheaper models?
At $5/$25 per million tokens, Opus 4.8 is 30-60× more expensive than Chinese alternatives like DeepSeek V4-Pro. If you need the absolute best coding quality and can afford it, Opus 4.8 is the clear leader. If cost matters more than the last few percentage points of quality, consider migrating to cheaper models.