Strongest long-context + multimodal blend
FRONTIER MODEL INDEX ↗
A practical field guide to the models shipping now — benchmark signals, context, modality, and where each one actually fits.
HOW TO READ THIS
Use the benchmark columns as signals, then use the fit notes to make a decision. No single score tells the whole story.
02 / THE WORKSPACE
| # | MODEL ↕ | RELEASE | CONTEXT | SWE-Procoding | Terminal 2.0agentic | GPQAscience | BrowseCompresearch | ARC-AGI 2abstract | BEST FIT | |
|---|---|---|---|---|---|---|---|---|---|---|
| 01 | G Gemini 3.1 ProGoogle · Closed | Feb 2026 | 1M | 54.2% | 68.5% | 94.3% | 85.9% | 77.1% | Strongest long-context + multimodal blend | |
| 02 | O GPT-5.5OpenAI · Closed | Apr 2026 | 400K | 58.6% | 82.7% | 93.6% | 84.4% | 85.0% | Best all-rounder for long-running work | |
| 03 | A Claude Opus 4.8Anthropic · Closed | May 2026 | 1M | — | — | 93.6% | — | — | Careful collaborator for complex agent work | |
| 04 | Llama 4 MaverickMeta · Open weights | Apr 2025 | 1M | — | — | — | — | — | Flexible multimodal base for custom products | |
| 05 | M Mistral Medium 3.5Mistral · Open weights | May 2026 | — | — | — | — | — | — | Self-hostable coding + agent workhorse | |
| 06 | x Grok 4.5xAI · Closed | Jul 2026 | — | 64.7% | 83.3% | — | — | — | Fast path for real-world engineering tasks | |
| 07 | D DeepSeek-V4 ProDeepSeek · Open weights | Apr 2026 | 1M | — | — | — | — | — | Open-weight frontier for self-hosted agents |
Strongest long-context + multimodal blend
Best all-rounder for long-running work
Careful collaborator for complex agent work
Flexible multimodal base for custom products
Self-hostable coding + agent workhorse
Fast path for real-world engineering tasks
Open-weight frontier for self-hosted agents
03 / SIGNALS, NOT SCORES
Public SWE-Pro results currently favor models tuned for autonomous software work.
At the frontier, the spread is tiny. Method and reasoning budget matter more than the decimal.
Context is a capability, not a guarantee. Retrieval quality and cost still decide production fit.