Anthropic shipped Claude Opus 4.8 on May 28, 2026. Same pricing as Opus 4.7. New default context window of one million tokens on the Claude API, Amazon Bedrock, and Vertex AI. Sharper judgment in agentic work. A new dynamic workflows tool in research preview for Claude Code that can coordinate hundreds of parallel subagents. And one number that is genuinely surprising: 4.8 is about four times less likely than 4.7 to let flaws in its own code pass without flagging them.
If you have been using 4.7, the upgrade is automatic on most plans. If you have not, this post is the no-fluff breakdown of what changed, what it costs, and whether to switch.
Quick facts. Launched: May 28, 2026, 41 days after Opus 4.7. API model name: claude-opus-4-8. Standard pricing: $5 per million input tokens, $25 per million output tokens. Fast mode: $10 in, $50 out. Context window: 1M tokens on Claude API, Bedrock, and Vertex AI; 200K on Microsoft Foundry. Max output: 128K on Messages API, 300K on Batch API beta. The headline shipping change is Dynamic Workflows in Claude Code, available on Enterprise, Team, and Max plans.
The honesty upgrade is the real story
Anthropic is leading with reliability rather than raw capability this release, and the headline number tells you why.
Opus 4.8 is roughly four times less likely than 4.7 to let flaws in the code it writes pass without flagging them. It is also more likely to surface uncertainty about its own work and less likely to make unsupported claims. Bridgewater Associates, an early customer, said the model's tendency to proactively flag issues with inputs and outputs to an analysis is something other models routinely missed.
That is the kind of upgrade that does not light up benchmark Twitter but quietly changes how much you have to review the model's output. If you are using Opus for production code or any agentic task where mistakes compound across many steps, this is the change that matters most.
Dynamic Workflows changes what Claude Code can attempt
The headline shipping feature is Dynamic Workflows for Claude Code, in research preview on Enterprise, Team, and Max plans. The pitch is simple. Instead of one agent doing everything in sequence, Opus 4.8 plans and runs hundreds of parallel subagents in a single session, then merges the results.
Anthropic's lead example is codebase-scale migration. The system can take a migration spec, fan out across hundreds of thousands of lines of code, use the existing test suite as the validation bar, and ship the work from kickoff to merge with a human in the loop only at decision points.
If that lands as promised, the practical change is not that Opus is smarter. It is that the unit of work you can hand to an AI in one shot just got an order of magnitude larger. Refactors that used to be a week of human-AI ping pong become a session.
Effort control is now a first-class UI
On claude.ai and Cowork, you now see an effort control next to the model selector. Opus 4.8 defaults to high effort, which means the model thinks longer before answering hard questions. You can dial it down for routine tasks where you want speed and lower cost.
Two practical implications. First, the same model can behave very differently depending on this knob, so reviews and benchmarks that do not name the effort setting are noise. Second, the cost of an Opus call is now partly a choice. Pick a low effort for short generation, high effort for planning and review.
Pricing is unchanged and that is the surprise
Same numbers as Opus 4.7. Five dollars per million input tokens, twenty five dollars per million output tokens on the standard tier. Ten and fifty on the fast tier. No price increase to absorb the better quality.
This is rare. Frontier models usually get more expensive as they get more capable. Pricing parity here makes Opus the easier swap for anyone already on 4.7, and a more honest comparison point for anyone considering it for the first time.
How 4.8 stacks up against 4.7 and GPT-5.5
Anthropic's own benchmark table shows 4.8 ahead of 4.7 across coding, reasoning, and knowledge work. It also reports 4.8 beating GPT-5.5 on several benchmarks at cost parity, and scoring 84 percent on Online-Mind2Web, the standard agentic computer-use benchmark.
Treat the headline numbers carefully. Vendor benchmarks favor the vendor. The honest test is to run your real workload on both models for a week and see which one ships more correct work per dollar. The pattern across daily use of 4.7 has been that the public scores understate how much smoother long agent loops feel in practice. If 4.8 holds the same pattern, the on-paper margin and the in-practice margin will not look the same.
What changed at a glance
| Opus 4.7 | Opus 4.8 | |
|---|---|---|
| Launched | April 16, 2026 | May 28, 2026 |
| API model name | claude-opus-4-7 | claude-opus-4-8 |
| Input price (standard) | $5 / 1M tokens | $5 / 1M tokens |
| Output price (standard) | $25 / 1M tokens | $25 / 1M tokens |
| Default context window | 200K | 1M (Claude API, Bedrock, Vertex) |
| Max output tokens | 128K (Messages API) | 128K (Messages API); 300K via Batch API beta |
| Effort control in UI | No | Yes, defaults to high |
| Dynamic Workflows | Not available | Research preview (Enterprise / Team / Max) |
| Code-flaw passthrough rate | Baseline | About 4x lower than 4.7 |
Why ship 4.8 only 41 days after 4.7
Forty-one days between Opus releases is fast even for Anthropic. The release notes do not say so directly, but the cadence is plausibly a response to two things: mixed user reception of some 4.7 behaviors in agentic settings, and competitive pressure from recent releases at OpenAI and Google.
The other context worth knowing: Anthropic announced $65B in new funding around the same window. A faster release cadence tracks with both the scale of capital flowing in and the urgency that flows out of it.
Should you switch from 4.7
On most plans the switch is automatic. The model name you call in the API changes from claude-opus-4-7 to claude-opus-4-8. The cost stays the same. The reliability gains on flagging uncertainty are immediate. The 1M default context window means you can stop chunking long inputs you used to chunk on 4.7.
Two things to test before you trust it for production. First, run your hardest current prompt through 4.8 with the same effort setting and see if the answer quality holds without surprises. Second, if you have a Claude Code workflow that does long migrations or refactors, try Dynamic Workflows on a non-critical project before you point it at your monorepo.
If you are not on Opus yet, the case is simpler. The honesty upgrade alone is reason to upgrade from Sonnet for any task where the model has to write code you will run.
The bottom line
Opus 4.8 is not a flashy release. The headlines are quieter than usual: same price, more honesty, longer context, a new workflow tool. None of that wins a press cycle. All of it changes the day-to-day math of building with AI.
If the four-times-fewer code flaws number holds up in real use, this is the release that makes Opus the default for production code agents, not just hard one-off prompts. That is a quieter but bigger shift than the version number suggests.
A follow-up after a full week of building with 4.8 on MemX is on the way. If you want to follow that thread, the product is at memx.app and the next post in this series will compare 4.7 and 4.8 head-to-head on the same set of real engineering tasks.
01When did Claude Opus 4.8 launch?
Anthropic released Claude Opus 4.8 on May 28, 2026, 41 days after Opus 4.7. It is available everywhere today via the Claude API using the model name claude-opus-4-8, and on claude.ai and Cowork for paid users.
02How much does Claude Opus 4.8 cost?
Pricing is unchanged from Opus 4.7. Standard: $5 per million input tokens and $25 per million output tokens. Fast mode: $10 per million input tokens and $50 per million output tokens. Anthropic publishes the live numbers on their API pricing page.
03What is the context window for Claude Opus 4.8?
One million tokens by default on the Claude API, Amazon Bedrock, and Vertex AI. Microsoft Foundry runs the 200K configuration. Max output on the Messages API is 128K tokens, with 300K available on the Batch API beta header.
04What is the difference between Claude Opus 4.7 and 4.8?
Opus 4.8 is roughly four times less likely than 4.7 to let flaws in its own code pass without flagging them, has a 1M default context window where 4.7 was 200K, ships with an effort control in the UI that defaults to high, and adds Dynamic Workflows in Claude Code. Pricing is identical.
05What are Dynamic Workflows in Claude Opus 4.8?
Dynamic Workflows is a research preview feature for Claude Code, available on Enterprise, Team, and Max plans. It lets Opus 4.8 plan and run hundreds of parallel subagents in a single session, designed for codebase-scale migrations that use the existing test suite as the validation bar.
