Claude Opus 4.8 Is Here: Faster, Cheaper Fast Mode, and Dynamic Workflows That Run Hundreds of Agents

Opus 4.8 builds on 4.7 with a focus on reliability and judgment over raw capability jumps. The most notable improvement, according to Anthropic, is honesty: the model is around four times less likely than Opus 4.7 to let flaws in code it has written pass unremarked. It flags uncertainty instead of plowing ahead with false confidence — a real quality-of-life upgrade for anyone using Claude Code on complex, multi-service work.

On benchmarks, Opus 4.8 is the only model to complete every case end-to-end on the Super-Agent benchmark — beating prior Opus models and GPT-5.5 at parity on cost. On CursorBench it exceeds prior Opus models at every effort level, with more efficient tool calling (fewer steps for the same intelligence). And on Online-Mind2Web — a browser-agent eval — it scores 84%, a meaningful jump over both Opus 4.7 and GPT-5.5.

The biggest practical unlock here is dynamic workflows, launching in general availability for Claude Code on Enterprise, Team, and Max plans. Claude can now plan a task and then spin up hundreds of parallel subagents to execute it — then verify outputs before reporting back. The example Anthropic gives: a full codebase migration across hundreds of thousands of lines of code, from kickoff to merge, with the existing test suite as its quality bar. That's a fundamentally different category of task than what Claude Code could handle before.

This is a direct answer to the scaling problem in large engineering organizations — not every task needs a human in the loop at each step. With Opus 4.8's improved judgment and honesty, it's actually credible to hand off something that big and trust the output. The fact that agents can now run longer within a session compounds this further.

Fast mode for Opus 4.8 — which lets the model work at 2.5x the speed — is now three times cheaper than it was for previous models. Pricing for regular usage stays the same as Opus 4.7: $5 per million input tokens, $25 per million output tokens. Fast mode drops to $10 input / $50 output per million tokens. For teams that were treating fast mode as a premium option, this changes the calculus significantly.

Beyond the model itself, Anthropic shipped a useful API update: the Messages API now accepts system entries inside the messages array. That means you can update Claude's instructions mid-task — changing permissions, token budgets, or environment context — without breaking the prompt cache or routing the update through a user turn. For anyone building complex agent harnesses, this removes a meaningful piece of friction.

The model is available via API today as claude-opus-4-8. Rate limits in Claude Code have been increased to accommodate higher token usage from elevated effort levels.

Anthropic was unusually direct about the roadmap. Two things are coming: models with the same Opus capabilities at lower cost, and a new class with higher intelligence than Opus. That second category is Mythos-class — currently in limited preview for cybersecurity work via Project Glasswing. Anthropic says they're making "swift progress" on the safety safeguards needed for general release and expect to bring Mythos to all customers "in the coming weeks."

If you're using Opus for production agentic work — coding, legal analysis, financial document processing, or anything long-running — Opus 4.8 is worth upgrading immediately. Same price, meaningfully better judgment, and the new dynamic workflows capability opens up a class of tasks that simply weren't feasible before. The honesty improvement alone (4x less likely to let code flaws slide) makes it a safer default for unsupervised Claude Code runs.

For teams debating whether to use Opus vs. Sonnet: the 3x cheaper fast mode and the new effort controls make Opus 4.8 more accessible at scale. If you've been holding off on Opus due to cost, this is a good moment to re-evaluate.

• Anthropic: Introducing Claude Opus 4.8