Three frontier releases at mid-tier prices

WEEK OF FEB 14 – 20, 2026

Three frontier releases at mid-tier prices

Anthropic, Alibaba and Google all shipped flagship-class capability inside four days — then a neutral benchmark reranked them and put cost per solve in charge.

Three flagship-class releases in four days, and all three landed at or below their predecessors' pricing. The capability race is now explicitly a price-performance race.

Two independent signals cut against the launch-day marketing: the first standardized SWE-bench leaderboard run put a last-generation model on top, and a well-known open-source agent author got hired by a frontier lab while his project was parked in a foundation.

THU · Feb 19, 2026ReasoningBenchmarks

Gemini 3.1 Pro previews a verified 77.1% on ARC-AGI-2

Google released Gemini 3.1 Pro in preview across the Gemini API, AI Studio, Gemini CLI, Antigravity, Android Studio, Vertex AI and Gemini Enterprise, and to consumers via the Gemini app and NotebookLM.

Google reports a verified 77.1% on ARC-AGI-2 — more than double Gemini 3 Pro's 31.1% — alongside 80.6% SWE-bench Verified, 94.3% GPQA Diamond and 69.2% on MCP Atlas, at the same price as Gemini 3 Pro, with a 1M-token input context.

WHY IT MATTERS

A roughly 2.3x jump on abstract reasoning in one point release is the largest single-generation gain of this cycle, and unchanged pricing puts frontier-tier reasoning at mid-tier cost.

Google blog Model card Simon Willison

THU · Feb 19, 2026EvaluationHarnesses

A neutral coding leaderboard puts a last-gen model on top

SWE-bench published a full official run of its Bash Only leaderboard, which drives every model through the same ~9,000-line mini-swe-bench agent with one shared system prompt instead of each lab's tuned scaffold.

The top of the board was Claude Opus 4.5 at 76.8%, then Gemini 3 Flash and MiniMax M2.5 tied at 75.8%, with Claude Opus 4.6 at 75.6%, GLM-5 and GPT-5.2 at 72.8%, and Kimi K2.5 at 71.4%. GPT-5.3-Codex did not appear. Three Chinese models made the top ten, and MiniMax M2.5 matched the leaders at roughly a twentieth of the cost per solved task.

WHY IT MATTERS

The rare coding benchmark where nobody grades their own homework — same harness, same prompt for every model. It makes cost-per-solve rather than raw score the decision variable for agentic coding.

Analysis SWE-bench leaderboards

TUE · Feb 17, 2026Frontier modelsPricing

Claude Sonnet 4.6 becomes the free-tier default

Anthropic released Claude Sonnet 4.6 as a full upgrade across coding, computer use, long-context reasoning, agent planning and knowledge work, with a 1M-token context window in beta, at unchanged pricing of $3/$15 per million tokens.

It is now the default for Free and Pro users in claude.ai and Claude Cowork. Reported results include 79.6% on SWE-bench Verified and 72.5% on OSWorld-Verified for computer use, with early-access developers preferring it over the previous-generation flagship Opus 4.5 59% of the time.

WHY IT MATTERS

It resets the free tier of a major assistant to something within a point of the paid flagship, which changes what baseline model quality means for anyone building on Claude.

Anthropic TechCrunch CNBC

MON · Feb 16, 2026Open weightsArchitecture

Qwen3.5 open-weights a 397B MoE that activates 17B

Alibaba's Qwen team released Qwen3.5-397B-A17B under Apache 2.0: natively multimodal with early fusion, 397B total parameters and only 17B activated per forward pass, built on a hybrid architecture fusing Gated Delta Networks linear attention with sparse MoE.

Language coverage expanded from 119 to 201, native context is 262K extensible to about 1M, and the team reports decoding throughput of 8.6x at 32K and 19x at 256K versus Qwen3-Max. Self-reported SWE-bench Verified is 76.4.

WHY IT MATTERS

The strongest evidence yet that the efficiency race — hybrid linear attention, extreme MoE sparsity, native multimodality — has replaced raw scale as the open-weight battleground.

Alibaba Cloud Model card Hugging Face analysis

MON · Feb 16, 2026PolicyInfrastructure

India hosts the first Global South AI summit

The India AI Impact Summit ran February 16–21 in New Delhi, the fourth in the global summit series and the first hosted by a Global South nation, drawing delegations from over 100 countries.

On February 19 tech minister Ashwini Vaishnaw announced the New Delhi Frontier AI Commitments, a voluntary framework adopted by leading frontier and Indian AI companies, plus a declaration endorsed by 92 countries. OpenAI used the event to launch OpenAI for India, citing over 100 million weekly ChatGPT users there and a Tata Group data-center partnership starting at 100MW and scaling toward 1GW.

WHY IT MATTERS

For developers the concrete news is capacity and distribution rather than the communiqué: sovereign data-center buildout, enterprise-wide agent deployments at a services giant, and a single country now accounting for a nine-figure weekly user base.

PIB India OpenAI for India

SUN · Feb 15, 2026AgentsOpen source

OpenAI hires OpenClaw's creator and parks the agent in a foundation

Sam Altman announced that Peter Steinberger, creator of the open-source personal agent OpenClaw, is joining OpenAI “to drive the next generation of personal agents.”

Steinberger said he could have built OpenClaw into a large company but found that unexciting, and that OpenClaw “will live in a foundation as an open source project that OpenAI will continue to support.” The hire followed weeks of courtship by multiple major labs.

WHY IT MATTERS

A clean test of whether agent value sits in the model or in the harness and integrations: a major lab bought the harness author, not the weights, and had to leave the agent open source to keep the community that made it work.

Reuters Steinberger's post TechCrunch