January was dominated by Chinese labs—Kimi K2.5, Qwen3-Max-Thinking, GLM-4.7-Flash—all shipping while Western labs were comparatively quiet. February looks different. Here's what's rumored, confirmed, and worth watching.
Confirmed
OpenAI Retires GPT-4o, GPT-4.1, o4-mini, and GPT-5 Variants (Feb 13)
OpenAI is forcing migration to GPT-5.2. On February 13, GPT-4o, GPT-4.1, o4-mini, and earlier GPT-5 variants go dark. If you're still pinning API calls to any of these, you have days to migrate. GPT-5.2 Codex also shipped in January for those on the coding-focused tier.
StepFun — Step-3.5-Flash (Feb 2)
StepFun quietly dropped Step-3.5-Flash on February 2. Another Chinese lab, another "Flash" variant optimized for speed. The pattern of lightweight, fast inference models from Chinese labs continues.
High Probability
Anthropic — Claude Sonnet 5 ("Fennec")
The most anticipated release this month. Dario Amodei's December 2025 podcast comments pointed to the next major release emphasizing "reliability" and "capabilities that matter for real work." Job postings reference "next-generation model deployment" and "safety testing for advanced capabilities."
A Vertex AI 404 screenshot surfaced showing claude-sonnet-5@20260203, but anyone can trigger 404s for nonexistent endpoints—not credible evidence on its own. Still, the timing lines up. Internal testing reportedly involves 1M token context windows.
Expected improvements:
The full Opus-class Claude 5 is further out—likely Q2-Q3.
DeepSeek V4 (~Feb 17)
The big one from China. The Information reports DeepSeek is targeting mid-February for V4, likely around February 17—timed with Lunar New Year, the same strategy they used with R1's January 2025 launch.
What we know:
GitHub updates revealed a new architecture identifier "MODEL1," widely interpreted as V4's technical foundation. If the performance claims hold, this is the release most likely to move markets—R1 triggered a $1 trillion tech sell-off last year.
Alibaba — Qwen 3.5 (~mid-Feb)
Alibaba is preparing Qwen 3.5, optimized for complex reasoning, math, and coding. This follows the rapid-fire Qwen3 series that dominated January (Qwen3-Max-Thinking topped "Humanity's Last Exam," Qwen3-Coder-Next matched DeepSeek-V3.2 on SWE-bench with 3B active params).
The timing is Lunar New Year—same window as DeepSeek. Alibaba recently connected Qwen to its e-commerce platforms, travel services, and Ant Group payments, positioning the chatbot as a comprehensive life assistant with 100M+ monthly active users. Qwen 3.5 likely powers that next step.
ByteDance — Doubao 2.0, Seeddream 5.0, Seeddance 2.0 (~mid-Feb)
ByteDance is going wide with three simultaneous launches:
ByteDance leads China's consumer AI space by user count. The multi-modal triple launch suggests they're betting on an integrated text-image-video pipeline rather than competing on LLM benchmarks alone.
Zhipu AI — GLM-5 (timing unclear, likely Q1)
Zhipu AI's founder Tang Jie announced GLM-5 in an internal letter on the day of their stock listing (January 8), promising it "will soon be unveiled" with "further scaling and numerous innovative technical improvements." Three stated technical directions: novel architecture design, a more universal RL paradigm, and continuous learning with autonomous evolution.
Context: GLM-4.7 already scored 84.9% on LiveCodeBench (beating Claude Sonnet 4.5) and 73.8% on SWE-bench Verified—highest among open-source models at the time. GLM-4.7-Flash dropped January 19 with a 30B MoE / ~3B active design. GLM-5 should be a significant step up.
Zhipu also shipped GLM-Image in January—notable because it was trained entirely on Huawei Ascend hardware, proving the viability of a fully domestic Chinese AI stack independent of NVIDIA.
Worth Watching
Meta — "Avocado"
Meta's AI unit is testing a frontier model code-named "Avocado" as the successor to the Llama series. With Meta planning $115-135B in AI capex for 2026, this is their most ambitious model yet. February release is possible but unconfirmed—could slip to March.
Anthropic Labs Division
Anthropic launched a "Labs" division led by Mike Krieger (Instagram co-founder). No product announcements yet, but the formation of this team suggests consumer-facing AI products are coming. Worth tracking for announcements this month.
Apple x Google — Siri + Gemini
Apple and Google announced a multi-year deal to power next-gen Siri with Google's 1.2T parameter Gemini model, running on Apple's Private Cloud Compute. The timeline for consumer rollout is "this year"—any February developer previews or API details would be significant.
The Bigger Picture
The story of February 2026 is consolidation and acceleration:
- Chinese labs are coordinating a Lunar New Year blitz. DeepSeek V4, Qwen 3.5, Doubao 2.0, and potentially GLM-5—all targeting the same mid-February window. This isn't coincidental; the holiday captive audience is a strategic launch window, and each lab is racing to define the narrative before the others.
- OpenAI is cleaning house. Retiring four model families at once is aggressive. They're betting everything on the GPT-5.2 line.
- Anthropic is overdue. Claude Opus 4.5 shipped in late 2025. A Sonnet 5 release would re-establish competitive positioning against GPT-5.2 and the Chinese frontier.
- MCP is the standard now. Anthropic's Model Context Protocol hit 97M monthly SDK downloads. Whatever models ship this month, MCP compatibility is table stakes for agentic use cases.
This post reflects publicly available rumors and analysis as of February 4, 2026. Nothing here is confirmed unless explicitly stated.