Anthropic releases Claude Opus 4.8 with Dynamic Workflows and a 3x cheaper fast mode

# Opus 4.8: Dynamic Workflows is the actually interesting bit

Anthropic shipped Claude Opus 4.8 on May 28, 41 days after Opus 4.7. The headline numbers are real: 88.6 on SWE-bench Verified, 69.2 on SWE-bench Pro, 1890 on GDPval-AA. Same $5 / $25 pricing as Opus 4.7. Fast mode runs at 2.5x output speed and is 3x cheaper than the previous Opus fast tier. Artificial Analysis's read is that 4.8 hits these scores with 15% fewer turns and 35% fewer output tokens per task than 4.7, which is the more interesting line on the page.

I am not particularly interested in the SWE-bench delta. Models will keep moving the SWE-bench number up until that benchmark stops being a useful proxy for production coding agents. We are most of the way there already. The number I read first on these announcement posts is now the per-task token count, not the headline score. Token count is a closer proxy for what a real team actually pays.

The part of the release worth a longer read is Dynamic Workflows. It is in research preview, attached to Claude Code, and the design is straightforward: Opus plans a job, fires off hundreds of parallel subagents in one session, then checks the results before handing them back. The stated example is a codebase-scale migration across hundreds of thousands of lines, with the existing test suite as the acceptance bar.

A few honest thoughts on that pattern, from the seat of someone who has shipped agent systems.

First, hundreds of parallel subagents is not new. Teams have been hand-rolling this since late 2024, with the same recurring pain points: result reconciliation, partial-failure handling, deduplication of writes to a shared filesystem, and the surprisingly hard problem of the subagent thinking it succeeded when the parent has no way to verify. The cost is not the model. The cost is the orchestration layer underneath, the one that has to take 200 partial results and decide what to do with them. Whatever Dynamic Workflows is doing under the hood is the engineering, not the model call.

Second, the framing of the existing test suite as the bar is the right framing. The hardest unsolved problem in agentic coding is not capability. It is: did the agent actually do the thing it said it did. Letting a test suite be the judge punts that question to the place it should have been all along, the code's own definition of working. If your test suite is thin, Dynamic Workflows will quietly write you a thinly correct migration.

Third, the most useful quote in the launch is buried. The Bridgewater testimonial says the biggest single difference in 4.8 over 4.7 is the tendency to proactively flag issues with the inputs and outputs of an analysis, something other models routinely missed and left to the users to catch. This is taste, not capability. It is the model saying your prompt is underspecified before producing the wrong thing, which is the single most underrated behavior in production LLM systems. If 4.8 has actually internalized this, the win is larger than the SWE-bench number suggests.

What I would do with this release, if I were on a team running Claude Code in anger this week:

- Pick one repo with a clean test suite and a real backlog of cross-cutting refactors. Run Dynamic Workflows against it. Measure two things: how many of the parallel subagents needed a human-in-the-loop intervention before merging, and the wall-clock from kickoff to merge versus your team's prior baseline. Those two numbers are the only ones that matter. - Do not migrate your default model to 4.8 in production yet. Run it side-by-side against 4.7 for two weeks on your top 200 production queries. Watch token counts, not just outputs. If 4.8 really uses 35% fewer output tokens, your invoice should drop visibly and immediately. - If you have a long-horizon agent that has been failing on the agent said it succeeded but did not, try the new model with no other changes. The honesty behavior, if it generalizes, fixes a class of bugs that no prompt patch ever did.

The pricing story is the small fix I have been waiting for. A 3x cheaper fast mode at $10 / $50 per million tokens makes Opus a plausible default for the medium tier of an AI stack, not just the hard-call tier. I have been telling portfolio CTOs for six months that they are overpaying frontier prices for routine calls. Fast Opus, at this price, is the first frontier model that almost makes the three-tier stack collapse into two. Almost.

A few things I will be watching for in the next two weeks:

- A real third-party Dynamic Workflows write-up. The first honest post-mortem on a migration that went sideways is going to teach more than the launch blog. If you ship one, send it to me. - Whether the honesty behavior holds outside coding. Bridgewater is one quote in a controlled context. I want to see it in customer-support agents, in voice, and in long-horizon ops loops before I cite it again. - Whether Mythos-class shipping in the coming weeks, Anthropic's words, not mine, is paired with concrete safeguard write-ups or just a marketing line. I am open to either reading. I would prefer the write-up.

For now: a real release, with a real number worth caring about (token count, not SWE-bench), and an orchestration primitive that is going to either save teams a quarter of work or quietly produce thinly correct migrations behind their backs. Both stories will be true. The interesting product question is which one is true at your company.

If you run Opus 4.8 against a real codebase this week, send me what you find.