On May 28, 2026 — just 41 days after Opus 4.7 — Anthropic released Claude Opus 4.8, its new flagship model. The rapid upgrade cycle reflects both competitive pressure from OpenAI (GPT-5.5, Codex) and Google (Gemini 3.5 Flash) and a market that demanded more after Opus 4.7's underwhelming reception.
Here's what changed, what it means for developers and enterprises, and why the 244-page system card might be the most interesting part of the release.
The Model: Modest Benchmarks, Real-World Gains
Opus 4.8 is available immediately on claude.ai, Claude Code, the API (claude-opus-4-8), and Cowork at the same standard pricing as Opus 4.7 ($5/M input, $25/M output tokens).
| Benchmark | Opus 4.8 | Opus 4.7 | Delta |
|---|---|---|---|
| SWE-bench Verified | 88.6% | 87.6% | +1.0 pp |
| SWE-bench Pro | 69.2% | 64.3% | +4.9 pp |
| Terminal-Bench 2.1 | 74.6% | 66.1% | +8.5 pp |
| Humanity's Last Exam (no tools) | 49.8% | 46.9% | +2.9 pp |
| OSWorld-Verified | 83.4% | 82.3% | +1.1 pp |
| Online-Mind2Web (computer use) | 84.0% | — | Best tested |
| Finance Agent v2 | 53.9% | 46.2% | +7.7 pp |
Dynamic Workflows: Orchestrating Hundreds of Parallel Subagents
The headline new feature is Dynamic Workflows — a research preview in Claude Code that lets Opus 4.8 plan complex tasks and spawn hundreds of parallel subagents in a single session.
Unlike traditional agent chaining or skill composition, Dynamic Workflows work differently:
- Claude writes a JavaScript orchestration script describing how to decompose and parallelize the task.
- A runtime executes the script in the background — the plan lives in code, not Claude's context window.
- Subagents report back and Claude verifies outputs before presenting results.
Anthropic demonstrates the capability with a vivid example: a codebase-scale migration across hundreds of thousands of lines of code, from kickoff to merge, using the existing test suite as the quality bar.
Dynamic Workflows are capped at 1,000 subagents and are available on Claude Code Enterprise, Team, and Max plans. A bundled /deep-research workflow ships as a built-in example.
3x Cheaper Fast Mode
Anthropic also slashed prices on Fast Mode — the accelerated inference path that produces tokens at roughly 2.5x normal speed:
| Tier | Input (per M tokens) | Output (per M tokens) |
|---|---|---|
| Opus 4.8 Standard | $5 | $25 |
| Opus 4.8 Fast Mode | $10 | $50 |
| Opus 4.7 Fast Mode (old) | $30 | $150 |
That's a 3x price reduction versus the equivalent tier on Opus 4.7.
Effort Control: Dial Thinking Depth Dynamically
A new effort control slider on claude.ai and Claude Cowork lets users choose how much thinking Claude invests per response:
- Higher effort: Deeper reasoning, better answers, higher token consumption
- Lower effort: Faster responses, slower rate-limit consumption
- Extra high (xhigh) and Max: For difficult tasks or long-running async workflows
Honesty & Alignment: Approaching Mythos Territory
Opus 4.8 is approximately four times less likely than its predecessor to let flaws in code it has written pass unremarked.
On the composite misalignment score (lower is better):
- Claude Opus 4.7: 2.5
- Claude Opus 4.8: 1.9
- Claude Mythos Preview: ~1.8
Opus 4.8 comes remarkably close to Mythos — the restricted model that Anthropic has kept behind closed doors since April 2026.
What This Means
- Multi-Agent Orchestration Goes Mainstream — Dynamic Workflows validate the thesis that the future is coordinated swarms, not a single super-model.
- Alignment Is a Competitive Differentiator — The 4x reduction in undetected code flaws is a concrete safety improvement enterprises care about.
- The Pricing War Is Accelerating — With Fast Mode dropping from $150/M to $50/M, the real competition is intelligence-per-dollar in autonomous workflows.
The Bottom Line
Claude Opus 4.8 isn't a generational leap — but it doesn't need to be. It's a disciplined, focused upgrade that addresses real pain points: model honesty, agent coordination at scale, inference cost, and user control.
Sources: Anthropic Blog, Claude Opus 4.8 System Card, TechCrunch
Top comments (0)