February 2026 AI Model Releases: Claude, Gemini, GLM-5 in Fierce Competition
February 2026 saw Anthropic, Google, and Zhipu AI launch new models. From Claude Opus 4.6 to Gemini 3.1 Pro to GLM-5, here's the analysis of this AI arms race.
February 2026 was genuinely packed. Anthropic, Google, and Zhipu AI all dropped new models around the Lunar New Year, making it one of the busiest months for AI model releases so far. Here’s what’s actually new and what it means for you.
Anthropic: Opus 4.6 and Sonnet 4.6
Anthropic fired two shots in early February—first the flagship, then the mid-tier. The release cadence is definitely accelerating.
Claude Opus 4.6
Released February 5, Opus 4.6 is the current top dog from Anthropic. Here’s what changed:
Context: 200K tokens baseline, 1M in beta. Working with large codebases feels smoother when you don’t hit the memory wall constantly.
Visible Thinking: Extended Thinking mode shows you the model’s reasoning process. Love it or hate it—at least it’s there.
Output Length: 128K tokens per response, the biggest in the industry. Useful if you’re generating massive amounts of content, though most users won’t need it.
Coding: SWE-bench Verified scores are solid. Cross-file refactoring and long-term planning definitely improved.
Claude Sonnet 4.6
Twelve days later (February 17), Anthropic dropped Sonnet 4.6. Pricing stayed the same—$3 input, $15 output per million tokens—but performance jumped noticeably. That’s rare.
Computer Use: OSWorld scores now approach human-level. Browser navigation, form filling, cross-tab coordination—these were AI weak spots, now they’re getting workable.
Enterprise Feedback: Rakuten AI called the iOS code “the best we’ve tested.” Specific. Sonnet is now the default on both claude.ai and Claude Cowork, so free users get access too.
Google: Gemini 3.1 Pro
February 19, Google shipped Gemini 3.1 Pro.
The positioning is clear: enterprise-grade reasoning. Google says this model is built for “problems where simple answers aren’t enough,” with notably better reasoning than 3 Pro.
Benchmarks: They claim ARC-AGI-2 score doubled versus the previous version. Google tends to push benchmarks hard—worth testing yourself.
SVG Generation: Can generate animated SVGs from text prompts directly. Practical for frontend devs and designers, might save some search time.
Industry Take: Gartner analyst William McKeon-White called it “good continued progress, not a fundamental game-changer.” University of Washington’s Chirag Shah reminded us that better reasoning doesn’t solve everything—AI still has other homework to do for complex tasks.
Chinese Contender: Zhipu AI GLM-5
Not to be outdone, China’s Zhipu AI released something worth attention.
Open-Source Leader: GLM-5 ranks #1 among open-source models on Artificial Analysis, with a Quality Index of 49.64, beating the previous leader Kimi K2.5.
Coding Performance: LiveCodeBench scores look strong—some items actually beat closed-source models.
Free Commercial Use: Open-source means free to use and deploy locally. Huge appeal for budget-conscious teams.
Side-by-side summary
| Model | Released | Type | Key facts from the vendor |
|---|---|---|---|
| Claude Opus 4.6 | Feb 5 | Proprietary | $5/$25 per million input/output tokens; 1M-token context in beta (Anthropic) |
| Claude Sonnet 4.6 | Feb 17 | Proprietary | $3/$15 per million tokens, unchanged from Sonnet 4.5; default model on claude.ai (Anthropic) |
| Gemini 3.1 Pro | Feb 19 | Proprietary, preview | 77.1% on ARC-AGI-2 versus 31.1% for Gemini 3 Pro (Google; model card) |
| GLM-5 | Feb 11 | Open weights, MIT licence | 744B-parameter MoE with about 40B active parameters (Z.ai; Hugging Face) |
Reading the benchmark claims
Each launch came with benchmark charts, and each vendor chose the tests where it looked best. A few points help when comparing them:
- Same test, same settings. Scores depend on reasoning effort, tool access and prompt scaffolding. A result with tools or at maximum effort is not comparable with one without.
- Vendor versus independent results. Google’s ARC-AGI-2 figure is labelled as verified by the ARC Prize in its model card, which carries more weight than a purely internal number. Independent leaderboards such as Artificial Analysis aggregate several tests but change as new models arrive.
- Preview status. Gemini 3.1 Pro launched in preview, so behaviour and limits can still change before general availability.
Open weights versus API models
GLM-5’s MIT licence means teams can download, modify and deploy it commercially. “Free” applies to the licence, not to running it: a 744B-parameter model, even with only about 40B parameters active per token, needs multi-GPU server hardware to self-host. For most small teams, the practical options are a hosted API from Z.ai or another provider, or a smaller open model. The trade-off against Claude or Gemini is control and data residency versus convenience and, often, top-end capability.
Why so many releases at once
February’s cluster reflects a faster release cycle across the industry: Anthropic shipped two models within 12 days, and Google followed with a mid-cycle “.1” update rather than waiting for a major version. For users, the practical effect is that model choice should be revisited regularly, and applications are easier to maintain if the model can be swapped through configuration rather than code changes.
What Should Developers Actually Pick?
Honestly, 2026 isn’t about “which AI is strongest” anymore. It’s about fit:
- Coding focus: Claude still leads
- Google ecosystem heavy users: Gemini integration is convenient
- Free open-source: GLM-5 is worth a look
Pricing is trending down. Sonnet 4.6 gave more for the same money, GLM-5 is free. The barrier to entry for smaller teams keeps dropping.
At the end of the day, there’s no “one AI to rule them all.” Picking the right tool for your specific task is more practical than chasing the latest and greatest.
Sources
Reader Dispatches
No reader responses yet. Share your technical perspective below.
Related Posts
Five Major AI Trends to Watch in 2026
From agentic AI to healthcare breakthroughs, from efficiency-first infrastructure to multimodal innovation, the AI industry in 2026 shows a shift from hype to practical value. This article synthesizes predictions from IBM, Microsoft, and Stanford HAI to analyze five key trends.
Anthropic and the Pentagon Negotiate Usage Limits on Claude in a $200M Contract
How Anthropic and the U.S. Department of Defense disagreed over usage terms for Claude, including autonomous weapons and domestic surveillance, ahead of a contract deadline, and what happened next.
U.S. Government Halts Use of Anthropic as OpenAI Signs Pentagon AI Agreement
In February 2026, the U.S. government moved to stop using Anthropic's products after a dispute over usage restrictions, and OpenAI signed an agreement with the Defense Department. What happened and what it means for AI governance.