Q1 2026 frontier models compared: what the vendor numbers do and do not tell you

Between February 5 and March 5, the three leading closed-model vendors each shipped a flagship or near-flagship: Anthropic's Claude Opus 4.6 and Sonnet 4.6, Google's Gemini 3.1 Pro, and OpenAI's GPT-5.4. This review covers those four releases, using only vendor pages and press coverage we could open, and is candid about what cannot be concluded from them.

The releases and their prices

Two observations. First, pricing has converged: the cheapest and dearest of these differ by 2.5x on input, not by an order of magnitude. Second, GPT-5.4 raised its per-token input price from GPT-5.2, which OpenAI says is offset by lower total token use ^[6]. That is plausible but is a claim about your workload, which only your own measurement can confirm.

What each vendor reported

Anthropic. Opus 4.6 scored 81.42% on SWE-bench Verified with a prompt modification, averaged over 25 trials, and Sonnet 4.6 scored 80.2% with a prompt modification ^[1][2]. Anthropic also lists effort controls (low, medium, high, max), adaptive thinking, a beta context-compaction feature and agent teams in Claude Code as a research preview ^[1].

Google. Gemini 3.1 Pro posted a verified 77.1% on ARC-AGI-2, which Google describes as more than double the reasoning performance of Gemini 3 Pro ^[3]. Anthropic's Sonnet 4.6 reported 60.4% on the same test at high effort ^[2].

OpenAI. GPT-5.4 scored 83% on GDPval, which measures professional tasks across 44 occupations, against 70.9% for GPT-5.2 ^[5][7]. It reached 75.0% on OSWorld-Verified, a test of operating a computer through screenshots, against 47.3% for its predecessor and a 72.4% human baseline ^[5]. OpenAI also reports that its tool-search feature cut token use by 47% in early testing with MCP tools without a reported accuracy loss ^[5][7].

Why the numbers do not line up

• Different tests. Each vendor foregrounds the benchmark it does best on. SWE-bench Verified, ARC-AGI-2, GDPval and OSWorld measure different things, and no single table here covers all four models.

• Different settings. Anthropic's SWE-bench scores use a prompt modification and many trials ^[1][2]; its ARC-AGI-2 figure for Opus was run at maximum effort with a large thinking budget ^[1]. Scores at different effort levels cannot be compared directly.

• Vendor-run. Most of these results were produced by the vendors themselves. Analysts quoted by Techstrong.ai called GPT-5.4 "an evolution rather than a revolution" and described benchmark leapfrogging among Google, Anthropic and OpenAI, with the industry trending toward feature parity ^[5].

What changed for agents

The common thread across the four releases is less raw intelligence than agent plumbing. OpenAI describes five reasoning-effort settings for GPT-5.4, from none to xhigh, with fine-grained sampling controls available only at "none" ^[6]. Anthropic's Opus 4.6 exposes four effort levels and lets the model decide when deeper reasoning is worthwhile ^[1]. Computer use is not unique to OpenAI either: Techstrong.ai notes Anthropic introduced it in late 2024 ^[5]. For buyers, this makes cost a dial you set per task, so a routing policy (cheap setting by default, expensive setting on request) matters as much as the model name.

Open-weight alternatives

The closed flagships are not the only story. Mistral released Small 4 on March 16 under Apache 2.0: a 119-billion-parameter mixture-of-experts model with about 6 billion active parameters per token, a 256k context window, and API pricing of $0.15 input and $0.60 output per million tokens ^[8]. It is not a flagship competitor, but the price gap shows why many routine workloads no longer need a top-tier model.

How to choose

1. Build a test set from your own work, 50 to 100 real examples, and score each candidate on it.

2. Compare cost per completed task, not cost per token. Longer reasoning and extra tool calls change the total.

3. Check feature fit: context behaviour on long documents, tool-calling reliability, and data terms.

4. Keep a second model wired in. Swapping vendors is far easier when your prompts and tools are not tied to one provider's interface.

For most businesses, the practical implication is that the choice among these four will be settled by measured results on their own tasks and by cost per task, and that designing for easy model switching is worth more than predicting which vendor leads next quarter.