Summary: Anthropic shipped Claude Opus 4.8 on May 28, 2026, keeping standard pricing while adding Dynamic Workflows, effort control, and mid-task system entries. Reported gains include SWE-bench Pro rising from 64.3% to 69.2%, with better calibration and lower unflagged-defect likelihood. The article treats the release as a shift toward reliable, cost-controlled long-task work, while urging harness-specific reproduction before benchmark conclusions.
Forty-one days after Opus 4.7, on May 28, 2026, Anthropic shipped Claude Opus 4.8, and the new model was live on the API and in every client that same day, with pricing untouched: $5 per million input tokens, $25 per million output tokens. The company’s own framing is blunt — this is its strongest coding model yet, and one that is “more likely to flag uncertainties about its work and less likely to make unsupported claims.”
The story for developers is less about raw capability and more about three changes to daily work: Dynamic Workflows, which coordinate hundreds of parallel subagents inside a single session; effort control, adjustable per task; and a Messages API that now accepts system entries partway through a conversation.
The 41-Day Turnaround
The interval is TechCrunch’s headline: 41 days between generations is much quicker than the Opus line’s past rhythm. Two pressures explain it — a lukewarm market reception for Opus 4.7, and rivals closing in through OpenAI’s Codex push and Google’s Gemini 3.5 Flash, which shipped on May 19. The standard tier keeps 4.7’s pricing, while the new Fast mode bills $10 per million input and $50 per million output, running roughly 2.5x faster at 3x less cost than the fast modes on earlier models. Anthropic has turned speed into an add-on with its own price tag.
Where the Scores Moved
- SWE-bench Pro climbs from 64.3% on 4.7 to 69.2%
- Online-Mind2web lands at 84%, clear of both Opus 4.7 and GPT-5.5
- Artificial Analysis counts 15% fewer turns and 35% fewer output tokens per task against 4.7 — though turns still run about 30% higher than GPT-5.5
- GDPval-AA tops the table at 1,890, a 137-point gain over 4.7
- The model is roughly one-fourth as likely as 4.7 to let defects in its own code slip past unflagged
A caveat belongs beside those numbers: GPT-5.5’s 83.4% on Terminal-Bench 2.1 came under the Codex CLI harness, and cross-model comparisons hinge heavily on the harness and evaluation conditions. Before committing, look for reproductions in a setup that resembles yours rather than trusting a single leaderboard row.
Dynamic Workflows: Hundreds of Subagents, One Session
Dynamic Workflows lands in Claude Code as a research preview for Enterprise, Team, and Max plans. The flow: Claude lays out a plan first, spins up hundreds of parallel subagents within a single session, then checks the results. Anthropic’s reference case is concrete — a migration spanning hundreds of thousands of lines across a whole codebase, taken from kickoff to merge, with the existing test suite as the pass bar. Teams that hand-roll an orchestrator for exactly this kind of job are watching the orchestration layer get absorbed into the model itself; what frameworks used to bolt on is becoming native model behavior.
Effort Dials and Mid-Task Instructions
Effort control is available across all plans on claude.ai and Cowork, set to “high” by default, with “extra” (“xhigh” in Claude Code) and “max” for genuinely hard problems — so the priciest reasoning no longer has to be spent on routine work. The Messages API change is small on the surface and big in practice: system entries can now sit inside the messages array, so instructions can change mid-task without invalidating the prompt cache. Cache reuse is where a long-running agent’s token budget actually lives, and a common workaround just became supported behavior.
Calibration Wins, Mythos Still Gated
Calibration is the improvement early testers name most. Bridgewater singled out the model’s habit of proactively raising issues with an analysis’s inputs and outputs. Scott Wu, CEO of Devin (Cognition), reports that 4.7’s verbose comments and tool-calling problems are fixed. Databricks measured token costs about 61% below 4.7 in its Genie environment, and one tester called it the first model to complete every case end-to-end on a “Super-Agent” benchmark. On alignment, misaligned behavior rates run substantially below 4.7 and sit near Claude Mythos Preview, with “new highs” on Anthropic’s measures of prosocial traits. Mythos remains gated — its preview raised cybersecurity concerns — and the official word is that strengthened safeguards will bring it to all customers “in the coming weeks.”
That matches what we argued in our 2026 opening outlook: generation gaps keep compressing, and the contest is shifting from single-turn capability to reliability and cost control over long tasks. Opus 4.8 is the clearest example of that shift yet.