AI This Week: Sonnet 4.6, Gemini 3.1 Pro, and seven models in one month

Seven frontier models in one month. That's the tally for February 2026 as I write this. The releases keep coming. The gap between open-source and proprietary is close to zero on most benchmarks now. Here's what happened this week.

Top stories this week

Sonnet 4.6 delivers Opus-class performance at Sonnet prices

Anthropic released Claude Sonnet 4.6 on February 17th. 79.6% on SWE-bench Verified, 72.5% on OSWorld. Opus 4.6, for comparison, sits at 80.8% and 72.7%. A 1.2 point gap on coding. Basically nothing on computer use.

The pricing is what makes this interesting. Sonnet 4.6 costs $3 per million input tokens and $15 per million output tokens, five times cheaper than Opus. It also ships with a 1 million token context window in beta, a first for any Sonnet-class model.

Anthropic made it the default model in claude.ai and Claude Cowork. In Claude Code testing, developers preferred Sonnet 4.6 over Sonnet 4.5 70% of the time, and over the previous flagship Opus 4.5 59% of the time. That last number should worry anyone selling premium-tier models. When the cheaper option wins most head-to-head comparisons against the pricier one, the pricing structure gets hard to defend.

I've been using it since launch. For most of my coding work, the quality gap with Opus 4.6 is imperceptible. The cost gap is not. This is the kind of release that changes what model you reach for by default.

Gemini 3.1 Pro doubles its reasoning scores

Google released Gemini 3.1 Pro on February 19th. The number that matters: 77.1% on ARC-AGI-2, a benchmark built to test whether a model can solve entirely new logic patterns it never saw during training. That's roughly double the previous Gemini 3 Pro's score, 24% ahead of GPT-5.2, 9% ahead of Opus 4.6.

On coding, it holds its own too: 80.6% on SWE-bench Verified, tied at the top of the leaderboard with Opus 4.6. Pricing holds at $2 per million input tokens and $12 per million output tokens for prompts under 200K tokens. Existing users get it at no extra cost.

The ARC-AGI-2 result is the one worth sitting with. Reasoning has been the weak spot for most models. Doubling performance on a benchmark specifically designed to resist memorization shortcuts points to a real architectural or training gain, not just more compute thrown at the problem.

Windsurf Wave 13 ships parallel multi-agent coding

Windsurf released Wave 13 with three notable additions: their SWE-1.5 model free for everyone through March, parallel multi-agent sessions via Git worktrees, and a multi-pane interface for watching several agents at once.

The Git worktrees integration is the real feature here. You can run five Cascade agents on five different bugs at the same time, each on its own branch, sharing Git history without stepping on each other's files. Each agent gets its own dedicated zsh terminal with environment variable inheritance.

It's Windsurf's answer to Claude Code's Agent Teams, which shipped with Opus 4.6 earlier this month. Different implementation, same idea: coding is starting to look more like running a team of agents than writing code yourself. The tools are converging on multi-agent orchestration as the default mode.

Giving away SWE-1.5 free for three months is a smart acquisition play. Windsurf is betting that once developers get used to the model, the switching cost keeps them around after the free period ends.

Claude Code Security crashes cybersecurity stocks

On February 20th, Anthropic announced Claude Code Security, a research preview that scans codebases for vulnerabilities and suggests patches for a human to review. Using Opus 4.6, Anthropic's team found over 500 vulnerabilities in production open-source codebases, bugs that had gone unnoticed despite years of expert review.

The market panicked. CrowdStrike dropped 7.8%, Cloudflare fell 5.9%, Palo Alto Networks declined 6.4%. Smaller static-analysis vendors like JFrog fell nearly 25%. Billions in market cap gone in a few hours.

The reaction was disproportionate. Claude Code Security does static analysis. It reads code and reasons about vulnerabilities. It doesn't test runtime behavior, can't send requests through an API stack, can't confirm whether a flaw is exploitable in a live environment. A code reviewer, not a penetration tester.

That matters, because most of the companies that got hit sell runtime security, endpoint protection, network monitoring. Claude Code Security doesn't compete with CrowdStrike's endpoint detection or Cloudflare's DDoS protection. It competes with Snyk, SonarQube, manual code audit. The market shot at the wrong target.

The signal is still real, though. An AI that reasons about vulnerability patterns instead of matching known signatures, that's genuine progress. Traditional SAST tools are going to have to evolve. But the cybersecurity industry as a whole isn't under threat from a code scanner.

The February rush: seven frontier models in one month

Step back for a second. Opus 4.6 on February 5th, GLM-5 on the 11th, Qwen3.5 on the 16th, Sonnet 4.6 on the 17th, Gemini 3.1 Pro on the 19th. DeepSeek V4 is still expected before the month is out. Seven frontier-class models in a single month.

Three things stand out.

The open-closed gap is close to zero. GLM-5 (MIT) hits 77.8% on SWE-bench. Qwen3.5 (Apache 2.0) costs 60% less than its predecessor. Sonnet 4.6, proprietary but cheap, delivers Opus-class performance at Sonnet pricing. The premium for proprietary frontier models is shrinking fast.

Prices are in free fall. When several frontier models compete for the same workloads, the price floor drops. Qwen3.5 at $0.18 per million input tokens, Sonnet 4.6 at $3, Gemini 3.1 Pro at $2. These numbers would have sounded absurd six months ago.

Model routing is becoming essential. No single model wins everything. Gemini 3.1 Pro leads on reasoning (ARC-AGI-2). Opus 4.6 leads on agentic coding (SWE-bench). GLM-5 offers the best cost-performance on Huawei hardware. Qwen3.5 dominates multilingual coverage. Teams that build routing logic to pick the right model per task get a structural cost advantage.

What does this actually change?

February 2026 will be remembered as the month frontier stopped being exclusive. A year ago, hitting frontier performance meant paying full price to one of three US companies. Today you can get there with an MIT-licensed Chinese model, an Apache 2.0 Chinese model, a European model hosted in Sweden, or a cheaper Anthropic model. The options exist.

For builders, the practical shift: stop hunting for "the best model" and start thinking in model portfolios. Route reasoning to Gemini 3.1 Pro, code to Opus 4.6, multilingual to Qwen3.5, cost-sensitive workloads to Sonnet 4.6 or GLM-5. Treat model choice like an infrastructure decision: the right tool for each job.

The cybersecurity market's reaction to Claude Code Security is a preview of what's coming. Every time AI gets credibly good at a professional task, the market for legacy tooling in that space gets repriced. Static analysis goes first. Runtime security, testing, monitoring will follow. None of it disappears, but tools that don't evolve will lose ground to the ones that do.

What to watch next week

• DeepSeek V4, expected any day now: Apache 2.0, trillion-parameter MoE, 1M context. If it delivers, that's another viable open-source option at the frontier level.
• Claude Code Security in practice: I'm testing it this week on my own projects. Curious how it handles real code, beyond the cherry-picked examples in the announcement.
• February's final tally: we should have a clear count of frontier releases by month's end. Seven or eight, it's already unprecedented. March is going to look quiet by comparison.