Anthropic Launches Opus 4.8 with Parallel Sub-Agent Workflows
Anthropic released Claude Opus 4.8 on May 28, 2026 — right on its two-month cadence — with two headline changes that matter beyond raw benchmark scores: a research-preview feature that orchestrates hundreds of parallel verifying sub-agents, and a Fast mode now priced 3× cheaper than before, while standard pricing stays flat.
What the Source Actually Says
The VC Corner's changelog confirms the key numbers: agentic coding rises to 69.2% (up from 64.3% on Opus 4.7, beating GPT-5.5 at 58.6%), and the model is 4× less likely than 4.7 to let its own code flaws pass unremarked. Anthropic frames this not as a raw capability bump but as a reliability improvement that compounds over long working sessions — the difference between an assistant you trust and one you audit.
Standard pricing is unchanged at $5/$25 per million input/output tokens. Fast mode is the clearest economic shift: repriced to $10/$50 per million at 2.5× speed, it cuts the previous fast-mode cost by two-thirds. Alex Albert (Anthropic) confirmed via X the routing logic: Fast for interactive, rapid-response work; standard Opus for longer async tasks. Toggle with /fast in Claude Code or join the API waitlist at claude.com/fast-mode.
Dynamic Workflows — arriving as a Claude Code research preview — is the structural change. The model writes an orchestration plan, launches hundreds of parallel sub-agents, runs adversarial verification across their results, and returns a single converged answer. YouTube coverage (Matthew Berman's 90-minute livestream) confirmed Dynamic Workflows was used to rewrite the Bun JavaScript runtime from Zig to Rust internally, hitting a 99.8% test-suite pass rate. The ultracode setting in Claude Code pairs maximum reasoning effort with Dynamic Workflows for the hardest batch tasks. Ethan Mollick independently documented Opus 4.8 autonomously building a full RPG — three PDF manuals, a playable adventure, and a deployed website — with zero human feedback, using Claude Code on Netlify.
A new effort dial (Low→Max, defaulting to High) lets users manage cost on simpler tasks. The 1M-token context window is on by default across the Claude API, Amazon Bedrock, and Vertex AI. Anthropic also signalled a "Mythos"-class model — higher intelligence than Opus, currently in limited cybersecurity preview — expected within weeks.
Strategic Take
Dynamic Workflows reframes the unit of agentic work: migrations, full test-suite generation, and codebase audits become single-instruction operations rather than team efforts. Fast mode's repricing makes high-frequency agentic iteration economically viable in production pipelines. The Mythos timeline — measured in weeks, not quarters — is the signal to watch next.
