Claude Opus 4.8 Tops GPT-5.5 With Dynamic Workflows and 4x Better Honesty

Anthropic shipped Claude Opus 4.8 on May 28, 2026 — just 41 days after Opus 4.7. That is the fastest turnaround for an Opus‑class model ever, signaling a strategy shift toward rapid incremental upgrades rather than monolithic annual launches, according to TechCrunch. The accelerated pace comes after Opus 4.7 received a chilly reception from users, and amid fresh competition from OpenAI Codex and Google Gemini Flash models.

The model ships everywhere — API, claude.ai, and Claude Code — at the same $5 per million input / $25 per million output pricing as Opus 4.7. The real news is under the hood: a new Dynamic Workflows tool that lets Claude coordinate swarms of subagents, a 4x improvement in honesty, and a 3x price cut on Fast Mode for teams that need speed.

The standout feature is Dynamic Workflows, a tool that lets Claude plan a large task, fan work out across tens to hundreds of parallel subagents, verify outputs, and report back with a complete result, as detailed by.¹ This is not just a model upgrade — it is a genuine architectural shift in how Claude Code operates.

According to early testers on Reddit, Claude now writes its own orchestration scripts, fans work out across multiple parallel subagents in a single session, and verifies outputs before returning results. One developer reported running up to 1,000 subagents in a single session — a scale previously impossible without custom orchestration code.

Opus 4.8 is roughly four times less likely than its predecessor to allow flaws in code it has written to pass unremarked, according to Anthropic's system card. The model had the lowest incorrect rate of six tested models on every factual benchmark, not because it answered more correctly, but because it abstained when uncertain.

Simon Willison called this approach "refreshing" in his review of the release, who highlighted Anthropic's own description of Opus 4.8 as "a modest but tangible improvement." According to Reuters, early testers found the model "more likely to flag uncertainties about its work and less likely to make unsupported claims."

Opus 4.8 took the #1 spot on the Artificial Analysis Intelligence Index at 61.4, edging past GPT‑5.5 at 60.2, according to Lushbinary. But the aggregate hides where each model actually wins.

While standard pricing stays at $5/$25 per million tokens, Fast Mode pricing dropped significantly. Opus 4.8 Fast Mode costs $10 per million input and $50 per million output — down from $30/$150 on Opus 4.7. That is a 3x reduction, making Fast Mode usable for development workflows, as Simon Willison detailed.

Fast Mode is currently limited to organizations in the research preview. But for teams that have it, the economics are now competitive with standard Opus pricing from months ago. Contra Collective notes Opus 4.8 remains roughly 4x more expensive on input than GPT‑5.5 ($5 vs ~$1.25 per million), but per‑task cost can favor Opus if it finishes in fewer turns.

Opus 4.8 ships several quality‑of‑life improvements. The most interesting is mid‑conversation system messages: developers can now inject updated instructions later in a long‑running conversation without restating the full system prompt, preserving prompt cache hits and reducing input costs on agentic loops, as Simon Willison documented.

The prompt cache minimum dropped from 4,096 to 1,024 tokens — shorter prompts now benefit from caching, a meaningful cost optimization for developers making frequent small API calls.

The 41‑day turnaround signals Anthropic is competing on iteration speed. This is the fifth Opus release in seven months. As Lushbinary noted, the two best generally available models are now separated by just 1.2 points on the aggregate index.

Looming behind Opus 4.8 is Claude Mythos, Anthropic's most powerful model with advanced cybersecurity capabilities, which Reuters reported will roll out to all customers "in the coming weeks." For builders, it represents the next frontier. For developers choosing today: if your workload is multi‑file code changes with surgical precision, Opus 4.8 is the best generally available model. For terminal‑driven command loops with lower input costs, GPT‑5.5 still has an edge — but the gap is narrowing fast.