The Frontier Model Race Heats Up: Claude Mythos Withheld, GPT-5.5 Incoming, and Gemini 3.1 Leads Benchmarks
- May 11
- 3 min read
The frontier model landscape is shifting faster than ever, and the latest round of releases shows a split between raw capability, controlled access, and task-specific leadership. Anthropic has disclosed Claude Mythos Preview through Project Glasswing rather than making it broadly available, OpenAI has now shipped GPT-5.5 after GPT-5.4’s March release, and Google’s Gemini 3.1 Pro is posting standout benchmark results. Here’s what it all means for builders.
Anthropic's Claude Mythos Preview: Restricted by Design
The biggest story this cycle is Anthropic’s decision to disclose Claude Mythos Preview through Project Glasswing rather than release it broadly through standard API access. Anthropic describes Mythos Preview as an unreleased frontier model being used with select partners to help secure critical software, citing cybersecurity risks as the reason for restricted access.
This is one of the clearest recent examples of a frontier lab restricting access to a highly capable model on safety grounds. For engineering teams, the practical implication is that frontier capability may increasingly be gated behind partnership agreements rather than made available through normal API access. Anthropic did release a smaller model alongside the announcement, but the main signal is the access model: the most capable systems may not always be the most accessible ones.
Meanwhile, the models you can use from Anthropic remain extremely strong. Claude Opus 4.6 holds the #1 spot on the LMSYS Chatbot Arena and hit 65.3% on SWE-bench Verified. Claude Sonnet 4.6 is the value play, performing at near-Opus levels at Sonnet pricing and leading the GDPval-AA Elo benchmark at 1,633 points. If you're building agents or code-generation pipelines, Sonnet 4.6 remains the best cost-performance ratio in the market.
GPT-5.4 Delivered, GPT-5.5 Has Arrived
OpenAI’s GPT-5.4, released in March, delivered strong benchmark results, particularly on computer-use tasks. Its scores on OSWorld-Verified and WebArena Verified made it a leading choice for browser automation and computer-use agent architectures, and its 83.0% score on OpenAI’s GDPval benchmark was also notable.
But the bigger update is that GPT-5.5 has now shipped. OpenAI says GPT-5.5 reaches 84.9% on GDPval, 78.7% on OSWorld-Verified, and 98.0% on Tau2-bench Telecom, making it a stronger candidate for complex professional workflows, coding, research, and agentic tasks.
For teams evaluating model choices, GPT-5.5 now resets the comparison point for OpenAI’s lineup. Claude remains highly competitive for coding and reasoning workflows, while Gemini 3.1 Pro is especially strong for multimodal and scientific reasoning. The gap between frontier models continues to narrow, which means model choice increasingly comes down to task fit rather than overall capability.
Gemini 3.1 Pro: The Quiet Benchmark Leader
Google's Gemini 3.1 Pro deserves more attention than it's getting. Posting standout results across several major benchmarks and hitting 94.3% on GPQA Diamond (a graduate-level reasoning benchmark), it's arguably the most well-rounded model available right now. The addition of real-time voice and image analysis makes it particularly strong for multimodal applications.
The Apple-Google partnership (covered separately) validates Gemini's capabilities. Apple chose a custom 1.2T parameter Gemini model over building their own frontier model, which speaks volumes about where Google DeepMind stands technically.
What This Means for Engineering Teams
The practical takeaways for AI engineering teams:
Multi-model architectures are now essential. No single model leads across all tasks. Route by capability: coding and reasoning tasks to Claude or GPT-5.5, computer-use workflows to GPT-5.5, and multimodal reasoning to Gemini.
Cost-performance optimization matters more than ever. Claude Sonnet 4.6 performing at near-Opus levels at a fraction of the cost is the kind of efficiency gain that changes your unit economics. Don't default to the most expensive model.
Plan for gated access. Anthropic's decision to restrict Mythos signals a trend. The most powerful models may not be available via standard API access. Build your architectures to be model-agnostic.
Benchmark on your actual tasks. GDPval, SWE-bench, and GPQA Diamond are useful directional signals, but the variance between models on your specific workload can be significant. Run your own evals.
The frontier is moving fast, and the gap between "best model" and "best model for your use case" is widening. Build accordingly.
