Anthropic's Claude Opus 4.8 Beats GPT-5.5 With a 67% Win Rate

Anthropic had a big day. The company released Opus 4.8, the newest version of its most advanced publicly available model, and simultaneously closed a $65 billion Series H round at a $965 billion post-money valuation. The model launch and the funding round together paint a picture of a company that has gone from safety-focused underdog to the most valuable AI startup in the world, all in the span of a few months.

A benchmark built for the real world

To understand why the benchmark numbers here matter, you need to understand what GDPval-AA actually tests. Most AI benchmarks are multiple-choice or short-answer tests. GDPval-AA is different. The benchmark requires models to produce diverse outputs including documents, slides, diagrams, and spreadsheets, mirroring actual work products across finance, healthcare, legal, and other professional domains. Models are dropped into a sandboxed environment with shell access and web browsing, then scored by an LLM judge doing blind pairwise comparisons of outputs from different models on the same task.

The GDPval gold public dataset includes 220 tasks developed by OpenAI in collaboration with industry professionals to reflect real-world complexity. Think: a retail supervisor creating a daily task list PDF, an A/V tech drafting a stage plot, a regional director building an Excel planogram tool. These are the kinds of deliverables that knowledge workers produce every day, not the kind of puzzles that show up in academic benchmarks.

The numbers that matter

Claude Opus 4.8 (Adaptive Reasoning, Max Effort) scores the highest on GDPval with a score of 1890, followed by GPT-5.5 (xhigh) with a score of 1769, and GPT-5.5 (high) with a score of 1753. That 121-point gap over GPT-5.5 translates to a roughly 67% head-to-head win rate against OpenAI's second-ranked model. The jump from Opus 4.7 is also notable: +137 Elo points in a single generation.

But the benchmark score is only part of the story. Efficiency improved dramatically too:

• Opus 4.8 completes tasks in 15% fewer turns than Opus 4.7
• It uses 35% fewer output tokens than its predecessor
• Opus 4.8's fast mode runs at 2.5x the speed, and is now three times cheaper than prior models
• Pricing starts at $5 per million input tokens and $25 per million output tokens, with up to 90% cost savings with prompt caching

One caveat worth noting: GPT-5.5 still leads on agentic terminal coding (Terminal-Bench 2.1), and Anthropic ran its Terminal-Bench scores through a different harness than OpenAI used. Benchmark methodology comparisons across labs always deserve scrutiny.

What actually changed under the hood

The new model comes just 41 days after Opus 4.7 was released, a much faster upgrade cycle than normal for Anthropic. The fast turnaround may have something to do with the chilly reception to Opus 4.7, which some users found disappointing. Whatever the motivation, the result is a model that feels qualitatively different in agentic settings.

The headline capability improvements across other benchmarks:

• Opus 4.8 scored 69.2% on SWE-Bench Pro, outperforming GPT-5.5 and Gemini 3.1 Pro
• Claude Opus 4.8 is the strongest computer-use and browser-agent model tested, scoring 84% on Online-Mind2Web, a meaningful jump over both Opus 4.7 and GPT-5.5
• On Anthropic's Super-Agent benchmark, Claude Opus 4.8 is the only model to complete every case end-to-end, beating prior Opus models and GPT-5.5 at parity on cost
• Claude Opus 4.8 delivers the highest score recorded on Anthropic's Legal Agent Benchmark, and is the first model to break 10% overall on the all-pass standard

Beyond raw scores, there is a more interesting behavioral shift. Early testers report that Opus 4.8 is more likely to flag uncertainties about its work and less likely to make unsupported claims. Evaluations show that Opus 4.8 is around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked. For anyone running long-horizon agentic pipelines, a model that surfaces its own uncertainty is worth more than a model that silently produces wrong outputs.

Dynamic Workflows: the feature builders should watch

Alongside the model itself, Anthropic launched a feature called Dynamic Workflows, available in research preview, designed to help larger models like Opus manage complex tasks across hundreds of parallel subagents. The mental model: an orchestrator session decides at runtime how many subagents to spawn, dispatches them in parallel, and synthesizes the results before handing them back.

Dynamic Workflows allows Claude to orchestrate up to 1,000 subagents running in parallel. Claude writes a JavaScript orchestration script, a runtime executes it in the background, and the system can handle codebase-scale tasks like code migrations across hundreds of thousands of lines. The feature is available for Claude Code for Enterprise, Team, and Max plans.

For developers, the migration path is minimal. The migration path is a one-line model ID change from claude-opus-4-7 to claude-opus-4-8. However, the new honesty-flagging behavior may attach short uncertainty notes to outputs that previously came back clean. Re-test downstream parsers before flipping production traffic.

The bigger picture: a near-trillion-dollar company

The model launch did not happen in isolation. The release came on the same day Anthropic closed a $65 billion Series H at a $965 billion valuation, ahead of a potential October 2026 IPO. The newest round almost triples Anthropic's valuation from February, when it was worth $380 billion. Anthropic's revenue has exploded thanks to its popular AI coding assistant, Claude Code. The company reported a $47 billion revenue run rate, up from a $30 billion run rate earlier this year, and $10 billion in annual revenue last year.

The $965 billion valuation puts Anthropic's financing above that of rival AI lab OpenAI. The competitive dynamic is now explicit: both Anthropic and OpenAI are racing toward public market debuts, and each new model release serves as both a product update and a demonstration of competitive strength for potential investors.

What comes next

The more intriguing storyline is what Anthropic is not releasing yet. Anthropic is still holding back its most advanced Mythos model after a tentative preview last month raised cybersecurity concerns. However, the company hinted in today's Opus release that the Mythos preview period might soon end, once necessary safeguards are complete. A small set of organizations already use Claude Mythos Preview for security work under Project Glasswing, but wider release waits on stronger safeguards.

For teams already running Claude in production, Opus 4.8 is a straightforward upgrade: same price, better performance, more honest about its own limitations. The official Anthropic announcement and the GDPval-AA leaderboard both have the full benchmark breakdowns. The model is available now via the Claude API as claude-opus-4-8, on AWS, Google Cloud, and Microsoft Foundry.