Claude Opus 4.8 Adds Effort Dials and a Token Meter

Anthropic released Claude Opus 4.8 on May 28, 2026, an upgrade to its flagship Opus 4.7 model that keeps the same token prices but adds user-adjustable effort controls, dynamic workflows in Claude Code, and live updates to the Messages API. The model is available now on claude.ai, the developer tool Claude Code, and the Claude API, where it is identified as claude-opus-4-8.

The benchmark gains are the easy headline. The change worth watching is quieter: a set of effort dials that let a single answer burn far more tokens than the last model did, landing on developers at the exact moment Anthropic is steering its whole business away from flat subscription tiers and toward a meter.

A Coding Upgrade Wrapped Around a Billing Shift

On the model itself, Anthropic makes the expected claims. Opus 4.8 beats its predecessor on coding, agentic tasks, reasoning, and office work, the company says in its Claude Opus 4.8 release notes. External testers cited in the announcement include firms in software, law, finance, and research. One coding evaluation, CursorBench, found Opus 4.8 needed fewer tool steps to reach the same output quality as comparable models, and Anthropic says one tester rated its cost performance on par with GPT-5.5 in internal benchmarks.

None of that is the structural news. The structural news is the effort control sitting next to the model picker. Users on claude.ai and the Cowork surface can now choose how hard Claude works on a response, and that choice maps directly to how many tokens the answer consumes. Opus 4.8 defaults to a high effort setting, with an xhigh tier above it for heavier jobs.

That dial only makes sense inside a billing model that charges per token rather than per seat. Anthropic has spent the year pushing customers toward usage-based pricing, and the effort slider is the consumer-facing edge of that move. The better the answer you want, the more the meter runs, and the user now holds the dial.

Effort Dials and the Token Meter

Opus 4.8 carries the same sticker price as Opus 4.7. Standard mode costs five dollars per million input tokens and twenty-five dollars per million output tokens. A faster mode runs at double those rates, and Anthropic says that fast mode now operates at 2.5 times the speed of standard while costing three times less than fast mode did on previous models.

The price card looks stable. What moves underneath it is consumption, and that is where the effort controls change the math. You can read the full rate card on the Claude API pricing schedule.

The Effort Levels

The slider gives users three positions, each spending more tokens than the last:

• High – the default, balancing quality against speed. On coding tasks, Anthropic says it spends a similar number of tokens as Opus 4.7’s default while delivering better results.
• Extra (called xhigh inside Claude Code) – more computation for harder problems.
• Max – maximum token spend for the best possible answer.

Why the Rate Limits Climbed

Because higher effort eats more tokens, Anthropic has raised Claude Code’s rate limits to keep the dials usable. That is the tell. A single edit command in Claude Code already bundles the system prompt, conversation history, file contents, and tool-use tokens into one call, so heavier effort settings can multiply a session’s bill quickly. Anthropic exposes spend controls and effort overrides through Claude Code’s cost-management settings, including the option to lower the effort level or cap the thinking budget.

Dynamic Workflows Target Codebase-Scale Jobs

The second new capability lives inside Claude Code. Dynamic workflows let Claude plan a job, fan it out across many parallel sub-agents, check its own work, then report back, all in one session. Anthropic says the feature is built for large codebases and can handle repositories running to hundreds of thousands of lines.

The loop runs in a fixed order:

1. Plan the work and break it into parallel tasks.
2. Launch hundreds of sub-agents that run concurrently, and with Opus 4.8, for longer stretches.
3. Verify the outputs against the existing test suite as the bar.
4. Report the merged result back to the developer.

The pitch is a codebase-scale migration carried from kickoff to merge with minimal hand-holding. The feature is in research preview and limited to Enterprise, Team, and Max plans, which keeps the heaviest token consumers inside the tiers most able to absorb a bigger bill. That placement is not an accident; the people most likely to run hundreds of agents at once are the ones already on metered enterprise contracts.

The Messages API Goes Live Mid-Task

For developers building on the API, the quieter change may matter more than the model scores. The Messages API now accepts real-time updates to the messages array while an agent is still working.

In practice, that means a developer can change instructions, adjust permissions, modify token limits, or refresh context partway through a long-running task. None of it requires a fresh user turn, and Anthropic says the updates do not break prompt caching, which is the mechanism that keeps repeated context from being re-billed at full rate.

It is a small spec change with a large operational effect. Long agent runs no longer have to be steered only at the start and inspected only at the end. They can be redirected while live, which is exactly what codebase-scale jobs need when something drifts off course halfway through.

Fewer Silent Failures in the Code

Anthropic frames the headline quality gain around a specific failure mode: code that looks fine and is not. The company says Opus 4.8 is markedly less likely than its predecessor to let its own mistakes slip through unflagged.

Around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked.

That line, from Anthropic’s release, is the most concrete reliability claim in the package. Alongside it, the company reports a cluster of safety and capability figures:

• 4x lower chance of passing flawed code without comment versus Opus 4.7.
• 10% cleared on the Legal Agent Benchmark’s all-pass standard, which Anthropic says makes Opus 4.8 the first model over that bar.
• Misaligned behavior rates that Anthropic describes as substantially lower than Opus 4.7, in line with its Mythos preview safety levels.
• Lower measured deception and a reduced tendency to comply with attempts to misuse the model.

The deception numbers connect to a longer thread of work. Anthropic has spent recent releases trying to wring out the kind of behavior it documented in earlier findings on Claude’s misaligned conduct, and the 4.8 alignment results read as the next step in that effort. For a coding tool sold on running unattended across a repository, fewer silent failures is the gain that actually justifies the higher token spend.

What the Six-Week Cadence Signals

Opus 4.8 arrived roughly six weeks after Opus 4.7, a release tempo that points at how Anthropic now sees the Opus line: not as a yearly event but as a steadily updated surface. The company says it is working on two fronts at once, models that deliver today’s capability more cheaply, and a new class that pushes past the current Opus ceiling.

That next class is Mythos. Anthropic says Mythos-class models, currently limited to participants in a preview it calls Project Glasswing, should reach customers in the coming weeks once additional safeguards are in place. The gating is deliberate, and Anthropic’s staged plan for releasing its Mythos-class models has already drawn attention to how much the safety review, rather than the training, now sets the calendar.

Read together, the pieces tell one story. Opus 4.8 hands users a faster, more careful coder and, in the same motion, a meter they control and are responsible for feeding. If the effort dials make sessions cheap enough to justify running hotter, the metered model wins and flat tiers fade out by default. If teams find the bill climbs faster than the quality, the dials become a cost-control chore, and the pressure swings back toward predictable pricing before Mythos even ships.

Frequently Asked Questions

Where can I use Claude Opus 4.8?

Opus 4.8 is available now on claude.ai, in the Claude Code developer tool, and through the Claude API, where it is called claude-opus-4-8. It launched on May 28, 2026.

Does Opus 4.8 cost more than Opus 4.7?

The per-token price is identical: five dollars per million input tokens and twenty-five dollars per million output tokens in standard mode, double that in fast mode. Your actual bill can still rise because higher effort settings consume more tokens per response.

What do the effort controls do?

They let you choose how hard Claude works on a response, across high, extra (xhigh in Claude Code), and max settings. Higher effort spends more tokens for better answers, so the control is effectively a quality-versus-cost dial.

Who can use dynamic workflows in Claude Code?

Dynamic workflows are in research preview and available on Enterprise, Team, and Max plans. They let Claude plan a job, run hundreds of parallel sub-agents, verify outputs, and report back across very large codebases.

What changed in the Messages API?

The Messages API now accepts live updates to the messages array while an agent is running. Developers can change instructions, permissions, token limits, or context mid-task without starting a new turn or breaking prompt caching.

When will Mythos-class models be available?

Anthropic says Mythos-class models, currently in the Project Glasswing preview, are expected to reach customers in the coming weeks, once further safeguards are in place.