AI Roundup February 2026: Gemini 3.1 Pro, Claude 4.6 Twins, and Grok Imagine Go Live

TL;DR

February 2026 was a pure frontier-model month. Google shipped Gemini 3.1 Pro and posted a 77.1 percent score on ARC-AGI-2, more than double what Gemini 3 Pro managed a quarter earlier. Anthropic answered with a two-model punch, Claude Opus 4.6 early in the month and Claude Sonnet 4.6 mid-month, the latter pushing near-Opus quality at Sonnet pricing. xAI rounded it out by taking Grok Imagine 1.0 into genuinely usable video and audio. If you build with this stuff, your default model probably changed in February.

Gemini 3.1 Pro Clears 77 Percent on ARC-AGI-2

On February 19, Google DeepMind released Gemini 3.1 Pro, and the headline number is the one everybody fixated on: 77.1 percent on ARC-AGI-2. ARC-AGI-2 is the benchmark specifically built so models cannot memorize their way to a high score. It hands you novel logic puzzles you have not seen in training and asks you to actually reason. Gemini 3 Pro, only about three months older, was sitting in the low thirties on the same test. Roughly doubling that in a point release is not normal cadence, and it is the part of this launch worth paying attention to.

Beyond the one stat, Gemini 3.1 Pro topped the majority of the benchmarks Google tracked, kept the 1M-token context window, and supports large output budgets in the tens of thousands of tokens. For builders, the practical read is simple: the hardest reasoning and long-context work you have been routing to a frontier model just got a meaningfully stronger default option, and it is the same Gemini API surface you already wired up.

The honest caveat: a benchmark win on abstract reasoning does not automatically translate to your specific agent loop or RAG pipeline. ARC-AGI-2 measures novel pattern-solving, not memory or tool discipline, and Gemini still forgets things mid-task like everything else. Treat 77 percent as a reason to re-run your own evals, not a reason to skip them.

Anthropic Ships the Opus 4.6 and Sonnet 4.6 Pair

Anthropic spent February answering with two releases instead of one. Claude Opus 4.6 landed early in the month as the flagship update, and brought a 1M-token context window into beta alongside the existing Sonnet tiers. Then on February 17, Claude Sonnet 4.6 arrived as the mid-tier workhorse, and this is the one most teams will actually feel.

The pitch on Sonnet 4.6 is the one builders care about: near-Opus quality at Sonnet pricing. That is the whole ballgame for anyone running Claude in production at volume. A mid-tier model that closes most of the gap to the flagship changes the math on agent fleets, long-running coding tasks, and anything where you were rationing Opus calls to keep the bill sane. The release also leaned into agent-team workflows and broader app surface, but the cost-per-quality story is the headline.

Why the two-model cadence matters

• Opus 4.6 for the hard ceiling: when you genuinely need the strongest reasoning and the 1M context, it is there.
• Sonnet 4.6 for the volume: ship the everyday agent and coding traffic here and keep the unit economics defensible.
• The takeaway: Anthropic is competing on the price-to-quality curve, not just the top of the leaderboard, and that is the axis that decides what you actually deploy.

xAI Takes Grok Imagine Into Real Video

On February 2, xAI released Grok Imagine 1.0, and unlike a lot of generative-video demos, this one shipped with the boring-but-important upgrades that make it usable: 10-second clips, 720p resolution, and substantially better audio. Generated video has been stuck in the territory of pretty-but-pointless for a while, and longer coherent clips with sound that does not fall apart is the kind of incremental jump that quietly crosses the line into something people will actually put in front of an audience.

xAI also pushed a Grok 4.20 beta in mid-February on the language side, keeping its release pace high. The interesting thing about Grok is positioning: it is bundled tightly into the X platform, so the distribution story is different from a standalone API. For builders, Grok Imagine is worth a look if you want fast, audio-aware short clips and you are not married to the open-weight stack. It is not the model you self-host, but it is a real option in the hosted video lineup now.

Key Takeaways

• Reasoning benchmarks moved fast: Gemini 3.1 Pro doubling its ARC-AGI-2 score in a single point release means the frontier is still gaining ground on genuinely novel problem-solving, not just memorized tests.
• Price-to-quality is the real battleground: Sonnet 4.6 delivering near-Opus output at Sonnet pricing is the kind of move that changes what you can afford to run at scale, which matters more than a leaderboard crown.
• Long context is now table stakes: both Gemini 3.1 Pro and Opus 4.6 carry 1M-token windows, so massive-context workflows are a baseline expectation across vendors, not a differentiator.
• Generated video crossed a usability line: Grok Imagine 1.0 hitting 10-second 720p clips with decent audio signals that hosted video gen is graduating from demo to deployable.
• Re-run your evals: three vendors shipped new defaults in one month, so whatever model your pipeline pinned in January is probably no longer the best pick, and the only way to know is to test on your own workload.