Just 41 days after shipping Opus 4.7 to mixed reviews, Anthropic is back with Claude Opus 4.8 — a frontier model upgrade that pairs benchmark gains in coding and agentic reasoning with a headline-grabbing new capability: Dynamic Workflows, a system that lets a single Claude Code session spawn up to 1,000 parallel subagents to tackle codebase-scale engineering tasks. The release, announced May 28, also slashes the price of the model\u2019s fast mode by a factor of three, a move that sharpens Anthropic\u2019s competitive positioning against OpenAI\u2019s Codex and Google\u2019s Gemini 3.5 Flash.
What Shipped
Opus 4.8 arrives at the same standard API pricing as its predecessor — \$5 per million input tokens and \$25 per million output tokens — but the economics shift meaningfully at the speed tier. Fast mode, which runs the model at roughly 2.5 times standard speed, now costs \$10 per million input tokens and \$50 per million output tokens, down from \$30 and \$150 respectively under Opus 4.7. For enterprise teams running high-volume agentic pipelines, that three-times cost reduction could reshape how aggressively they deploy Opus for latency-sensitive workloads.
On benchmarks, the gains are incremental but consistent. Opus 4.8 scores 88.6 percent on SWE-bench Verified, up from 87.6 percent for Opus 4.7, and 69.2 percent on the harder SWE-bench Pro, a nearly five-point jump from 64.3 percent. Terminal-Bench 2.1 saw a larger leap, from 66.1 percent to 74.6 percent. Anthropic says the model beats OpenAI\u2019s GPT-5.5 across at least 12 benchmarks spanning coding, agentic tool use, and long-context tasks.
The standout feature, however, is Dynamic Workflows. Available in research preview for Claude Code users on Enterprise, Team, and Max plans, the system allows Claude to decompose a large task, spin up hundreds of parallel subagents, execute work across them, and verify outputs before reporting back. Anthropic says Claude Code with Opus 4.8 can now perform codebase-scale migrations across hundreds of thousands of lines of code, running from kickoff to merge using an existing test suite as its quality bar. A single run can spawn up to 1,000 agents, though Anthropic cautions that costs can climb rapidly at that scale.
What the Industry Is Saying
Scott Wu, CEO of Devin, said in an early tester testimonial that Opus 4.8 \u201cuses tools cleanly and follows instructions with the consistency our autonomous engineering workloads need to keep running unattended.\u201d Wu added that the release \u201ctranslates directly into faster capability gains for engineers building on Devin.\u201d
Michael Truell, co-founder and CEO of Cursor, reported that on CursorBench, Opus 4.8 \u201cexceeds prior Opus models across every effort level\u201d and that tool calling is \u201cmeaningfully more efficient, using fewer steps for the same intelligence.\u201d
Hanlin Tang, CTO of Neural Networks at Databricks, called the model \u201ca step change in agentic reasoning\u201d and noted a 61 percent cheaper token cost compared to Opus 4.7 for multimodal workloads running through Databricks\u2019 Genie agent.
Alignment and Honesty
Anthropic\u2019s internal alignment assessment may be as significant as the performance numbers. The company\u2019s Alignment team concluded that Opus 4.8 \u201creaches new highs on our measures of prosocial traits like supporting user autonomy and acting in the user\u2019s best interest.\u201d Misalignment rates — covering behaviors such as deception or cooperation with misuse — are now on par with Claude Mythos Preview, the company\u2019s most advanced and still-restricted model. It is the first time a generally available Claude model has reached Mythos-class alignment levels.
On a practical level, Opus 4.8 is reported to be four times less likely than Opus 4.7 to allow flaws in code it produces to pass unremarked, and less prone to making unsupported claims about its own progress during agentic tasks. The company frames this as a step toward more honest AI collaborators, though it also flagged what it called \u201ca concerning trend that could complicate training in the future\u201d without elaborating.
Why This Matters
The 41-day turnaround from Opus 4.7 to 4.8 signals something beyond routine iteration. TechCrunch\u2019s Russell Brandom noted that the fast cycle likely reflects the \u201cchilly reception\u201d to Opus 4.7, which some developers found underwhelming. It also reflects intensifying competitive pressure: OpenAI launched expanded Codex capabilities earlier in May, and Google shipped Gemini 3.5 Flash with an explicit bet on agents over chatbots. Anthropic needed to respond, and it did so with both a model refresh and a genuinely new product surface in Dynamic Workflows.
The pricing move on fast mode is equally strategic. At \$10/\$50 per million tokens, Anthropic is making speed-sensitive agentic use cases dramatically more accessible, precisely the territory where OpenAI\u2019s Codex and Google\u2019s agent-first Gemini Flash are competing hardest. The three-times price cut reframes the cost calculus for any enterprise evaluating which frontier model to build its agent infrastructure on.
Dynamic Workflows also marks an architectural shift in how AI companies think about model deployment. Rather than simply offering a smarter model, Anthropic is packaging orchestration — the ability to plan, parallelize, and verify — as a first-party feature. This moves the company further from selling raw intelligence and closer to selling complete engineering workflows, a positioning that could prove durable even as raw model capabilities converge across providers.
What to Watch
Anthropic hinted in the Opus 4.8 announcement that the restricted Mythos model may soon be generally available. The company wrote that it is \u201cmaking swift progress\u201d on the cybersecurity safeguards required for broad release and expects to bring Mythos-class models to all customers \u201cin the coming weeks.\u201d If that timeline holds, Opus 4.8 may be remembered less as a destination than as the bridge between Anthropic\u2019s current frontier and its next one. Meanwhile, the real test for Dynamic Workflows will be whether enterprise teams adopt 1,000-agent runs in production — and whether the economics hold up when they do.
“Uses tools cleanly and follows instructions with the consistency our autonomous engineering workloads need.”— Scott Wu, CEO, Devin