1Executive Summary

February 2026 produced the most concentrated frontier model release window observed to date. Anthropic shipped Claude Opus 4.6 on the 5th and Claude Sonnet 4.6 on the 17th. Google released Gemini 3.1 Pro on the 19th. xAI released Grok Imagine 1.0 on the 2nd, OpenAI shipped GPT-5.3-Codex on the 5th, and Zhipu AI released GLM-5 on the 11th. Across the month, more than ten open-weight or proprietary frontier-class models reached availability — a pace independent observers characterized as the densest model release window in AI history.

The 1-million-token context window became the standard for frontier models. Claude Opus 4.6 and Sonnet 4.6 both shipped with 1M context at standard pricing without beta headers. Anthropic's Opus 4.6 set the highest task-completion time horizon estimated by METR — 14 hours and 30 minutes at the 50% mark — and the highest reported MRCR v2 score at 1M tokens (78.3%). Gemini 3.1 Pro doubled its predecessor's ARC-AGI-2 score (31.1% to 77.1%) and reached 94.3% on GPQA Diamond, the highest score reported on that benchmark at the time.

Anthropic raised $30 billion in a Series G funding round on February 12 at a $380 billion post-money valuation — the second-largest private financing round in technology history, behind only OpenAI's $40 billion raise of 2025. The round was led by GIC and Coatue, with co-leads from D. E. Shaw Ventures, Dragoneer, Founders Fund, ICONIQ, and MGX. Anthropic disclosed run-rate revenue of $14 billion at the time of the round, with Claude Code at $2.5 billion annualized and weekly active users having doubled since January.

To watch in March: OpenAI's expected release cadence step (GPT-5.4 timeline), the second draft of the EU Code of Practice on Marking and Labelling of AI-generated Content, Q1 venture deployment closing data, and continued open-weight competition from Alibaba, Zhipu, and DeepSeek.

2Large Language Models

February’s release cadence was tighter than any month previously recorded. The release windows of OpenAI, Anthropic, Google, xAI, Alibaba, Zhipu, and ByteDance overlapped within hours of each other on multiple occasions.

2.1Releases and updates

Anthropic released Claude Opus 4.6 on February 5, available immediately through claude.ai, the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry on Azure. Opus 4.6 ships with a 1-million-token input context and up to 128K output tokens. The model introduced agent teams — multiple Claude instances running in parallel with peer-to-peer communication via what Anthropic called the Mailbox Protocol — and context compaction, which automatically summarizes earlier portions of a conversation as the context window approaches its limit. METR's task-completion time horizon estimate places Opus 4.6 at 14 hours and 30 minutes at the 50% mark and 1 hour 3 minutes at the 80% mark, the highest reported among frontier models. Long-context retrieval on MRCR v2 at 1M tokens reached 78.3%, ahead of GPT-5.4 at 36.6% and Gemini 3.1 Pro at 18.3%.

Anthropic followed with Claude Sonnet 4.6 on February 17. Pricing of $3 per million input tokens and $15 per million output tokens, with 1-million-token context at standard pricing, made it the lowest-priced model with that context length on the market at the time. Anthropic reported that early-access users preferred Sonnet 4.6 to Sonnet 4.5 in approximately 70% of head-to-head tests, and to the prior flagship Opus 4.5 in 59% of tests. Sonnet 4.6 leads on GDPval-AA, an evaluation of economically valuable professional tasks, with an Elo score of 1633 against Opus 4.6 at 1606 and Gemini 3.1 Pro at 1317.

Google released Gemini 3.1 Pro on February 19 in preview through the Gemini API, Google AI Studio, Vertex AI, Gemini CLI, Android Studio, and the Gemini app for Pro and Ultra subscribers. Gemini 3.1 Pro is the first version-point increment in the Gemini 3 series and integrates the reasoning techniques first deployed in Gemini 3 Deep Think. The model scored 77.1% on ARC-AGI-2 — more than double the 31.1% recorded by Gemini 3 Pro three months earlier — and 94.3% on GPQA Diamond, the highest score reported on that benchmark at the time of release. Pricing came in at $2 per million input tokens and $12 per million output tokens, undercutting both Claude Opus 4.6 and GPT-5.3-Codex. The model supports a 1-million-token context window with multimodal input across text, image, audio, and video.

OpenAI shipped GPT-5.3-Codex on February 5, the first OpenAI release dedicated to coding workflows under the Codex line. GPT-5.3-Codex remained available alongside GPT-5.2 Instant and the general GPT-5.3 variants. The model targeted SWE-bench-class evaluations and reached the upper tier of leaderboards on terminal coding tasks. Gemini 3 Deep Think, released the week of February 12, was used by Google to solve problems in science, research, and engineering, including the disproof of a decade-old mathematics conjecture.

xAI released Grok Imagine 1.0 on February 2, providing a publicly documented multimodal generation API for both image and video output. xAI followed with Grok 4.2 Beta on approximately February 17, ahead of the full Grok 4.20 Beta 2 release in March. Grok 4.2 Beta retained the four-agent multi-agent architecture with the Grok coordinator and three specialised sub-agents.

February also produced unusually intense open-weight ecosystem activity. Zhipu AI released GLM-5 on February 11, a 744-billion-parameter Mixture of Experts model. ByteDance shipped Kimi K2.5 with reasoning capabilities and Doubao 2.0. MiniMax launched M2.5 and M2.5 Lightning on February 12, posting strong SWE-bench scores at low cost as open models. DeepSeek released V3.2. Alibaba's Qwen 3.5 family began rolling out mid-month with an agent-focused architecture.

2.2Benchmarks and capabilities

Three benchmarks moved meaningfully during February. ARC-AGI-2 — designed to test novel-pattern recognition that resists training-set memorization — doubled at the frontier from Gemini 3 Pro's 31.1% to Gemini 3.1 Pro's 77.1%. GPQA Diamond, a graduate-level science test, reached 94.3% on Gemini 3.1 Pro — the highest score reported on the benchmark to that date. The 1-million-token context retrieval benchmark MRCR v2 reached 78.3% on Claude Opus 4.6 at 1M tokens, against 36.6% on GPT-5.4 and 18.3% on Gemini 3.1 Pro at the same context length.

The LMSys Chatbot Arena, which reflects crowd-sourced user preferences across diverse tasks, registered Gemini 3.1 Pro Preview at 1500 Elo and Claude Opus 4.6 variants at 1504 Elo — effectively tied at the top during the month.

Long task-completion horizons emerged as a meaningful axis of differentiation. METR's estimate places Opus 4.6 at 14 hours and 30 minutes for 50% completion. The metric matters because the practical limit of agentic deployment is task length — a model that loses coherence after two hours cannot complete an eight-hour analytical workflow no matter how high its single-step benchmark scores.

Computer use approached the human-expert baseline. Sonnet 4.6’s improvement on OSWorld-Verified, combined with its prompt-injection resistance now on par with Opus 4.6, made browser-based and desktop-based agentic deployment more practical at lower cost than the Opus tier required.

2.3Pricing and access

Claude Opus 4.6 entered at $5 per million input tokens and $25 per million output tokens, with up to 128K output tokens per request. Anthropic introduced a Fast Mode for Opus 4.6 delivering up to 2.5x faster output at premium pricing of $30 per million input and $150 per million output — same intelligence, faster turnaround.

Claude Sonnet 4.6 entered at $3 per million input and $15 per million output tokens. Both Opus 4.6 and Sonnet 4.6 included 1-million-token context at standard pricing without beta headers. The previously charged premium for requests exceeding 200K tokens was retained at the standard rate but the long-context surcharge would later be removed entirely on March 13.

Gemini 3.1 Pro at $2 per million input and $12 per million output tokens undercut both Claude Opus 4.6 and OpenAI’s GPT-5.x line on per-token pricing while delivering competitive benchmark scores. Open-weight options at near-zero licence cost — Zhipu GLM-5, MiniMax M2.5, ByteDance Kimi K2.5, Alibaba Qwen 3.5, DeepSeek V3.2 — expanded the cost floor downward across the segment.

3Generative Media

Generative media activity in February concentrated in video, with two new flagship architectures and one new entrant in the space, alongside continued capability advances in image and audio.

3.1Image

xAI released Grok Imagine 1.0 on February 2, the company’s first multimodal generation product. The release covered both image and video output through a publicly documented API and partner platforms, allowing third-party integration into product pipelines without building a separate media stack. Adobe, Black Forest Labs, Stability AI, and Midjourney did not release new flagship image models during the month.

3.2Video

Kling 3.0, from Kuaishou, was released during February with multi-shot sequence support of three to fifteen seconds and subject consistency across different camera angles — a meaningful technical step that simplified narrative video workflows previously requiring manual frame stitching. Multi-character native audio with voice reference allowed consistent voices across cuts.

ByteDance Seedance 2.0 was released during February as the first AI video model with unified audio-visual joint generation. Generation produced sound and image simultaneously rather than as a post-processed audio overlay. The model supported twelve-file multimodal input, phoneme-level lip-sync in eight languages, and identity lock for character consistency across scenes. Both Kling 3.0 and Seedance 2.0 reached broader API distribution through FAL.AI and partner platforms.

Sora 2 and Sora 2 Pro continued in service from OpenAI — OpenAI's announcement of the Sora discontinuation would not arrive until March 24.

3.3Music and audio

Suno and Udio did not release new flagship music models during February; both continued under their existing version lines (Suno 5.x, Udio constrained by the October 2025 Universal Music Group settlement). ElevenLabs continued integration partnerships with video generation pipelines. Google’s ProducerAI, built on the Lyria 3 architecture released in late 2025, expanded inside Google Labs as a music generation tool for Workspace.

4Coding and Developer Tools

February's coding tools activity was driven by the Anthropic Series G announcement and the OpenAI GPT-5.3-Codex release. Both events underscored the segment's emergence as the largest revenue category in commercial AI.

4.1Releases

Anthropic launched Claude Code Security during February, an agentic feature that reviews codebases to identify security vulnerabilities. The feature operates inside the existing Claude Code product and was made available to existing subscribers. Claude Opus 4.6, released February 5, became the default coding model in Claude Code; the agent teams capability allowed multiple Claude instances to work in parallel on different parts of a project.

OpenAI released GPT-5.3-Codex on February 5 as a coding-specialised model under the Codex line. The release continued the pattern of separate Codex variants for coding workflows, a structure that GPT-5.4 would consolidate one month later.

4.2Pricing and licensing

Claude Code revenue at $2.5 billion annualized was disclosed during the Anthropic Series G announcement on February 12, more than double the level at the start of the year and an indicator of how rapidly coding-tool revenue compounded. Claude Code business subscriptions had quadrupled since the start of 2026, and enterprise users represented over half of Claude Code revenue. Cursor remained the largest standalone IDE-based AI coding product, while GitHub Copilot continued at scale.

Devin pricing remained at $20 Core plus $2.25 per Agent Compute Unit, the structure introduced in late 2025. The shift from the original $500 monthly enterprise tier broadened access materially.

4.3Adoption signals

Anthropic's disclosed Claude Code metrics during the Series G announcement provided the clearest single set of adoption data points of the quarter. Claude Code authored approximately 4% of all GitHub public commits worldwide — double the percentage from one month prior. Weekly active users had doubled since January 1. Eight of the Fortune 10 were now Claude customers, and customers spending more than $100,000 annually had grown sevenfold over the past year. Customers spending more than $1 million annually exceeded 500, against a dozen two years earlier.

The data point most relevant to the broader category was that 79% of OpenAI users also paid for Anthropic, per Ramp aggregated payment data. Coding-tool adoption was therefore additive rather than substitutional within the developer market. One in five businesses using Ramp paid for Anthropic, against one in twenty-five a year earlier.

5Industry Trends

February closed the third month of the most concentrated AI capital deployment cycle on record. The Anthropic Series G alone matched the average annual venture deployment of pre-2020 cycles. Concurrently, the EU AI Act's August 2 enforcement deadline drew nearer with the Digital Omnibus negotiation continuing in Brussels.

5.1Funding and valuations

Anthropic's Series G was announced on February 12. The $30 billion raise valued the company at $380 billion post-money, more than doubling the September 2025 Series F valuation of $183 billion. The round was led by GIC and Coatue, with co-leads from D. E. Shaw Ventures, Dragoneer, Founders Fund, ICONIQ, and MGX. Other significant investors included Accel, General Catalyst, Jane Street, Microsoft, Nvidia, and the Qatar Investment Authority. The round included portions of the previously announced Microsoft (up to $5 billion) and Nvidia (up to $10 billion) commitments from November 2025.

Anthropic disclosed financial data alongside the round. Run-rate revenue had reached $14 billion, growing more than tenfold annually in each of the prior three years since the company first earned revenue. Dario Amodei told CNBC that approximately 80% of Anthropic's business came from enterprises. The Series G is the second-largest private technology financing round in history, behind OpenAI's $40 billion raise of 2025.

OpenAI itself was reportedly seeking an additional $100 billion in February, in talks that would close at $122 billion at the end of March. Other significant Q1 closings included xAI at $20 billion, Waymo at $16 billion, Databricks at $7 billion, Polymarket at $2.6 billion, and Shield AI at $2.3 billion. Mistral, while not closing a round during the quarter, hit $400 million in annualized recurring revenue per CEO Arthur Mensch's public statement.

5.2Regulation and legal

The EU AI Act 2 August 2026 enforcement deadline remained in force at the close of the month. The Digital Omnibus on AI — adopted by the European Commission in November 2025 — continued through Parliament and Council negotiation, with a political agreement targeted for March. The Digital Omnibus proposed a maximum postponement to 2 December 2027 for certain high-risk system obligations under Annex III, while preserving the original 2 August 2026 schedule for Article 50 transparency obligations.

US state-level regulation continued to fill the federal gap. The Texas Responsible Artificial Intelligence Governance Act, which had taken effect on January 1, 2026, entered its first month of practical enforcement. The Utah Artificial Intelligence Policy Act applied disclosure requirements to deployers of generative AI in regulated and consumer transactions. The Colorado AI Act, due to apply in June, drove pre-compliance preparation across enterprise deployments.

The principal generative AI copyright cases — New York Times v. OpenAI in the Southern District of New York and Getty Images v. Stability AI in the United Kingdom and the United States — continued through pre-trial motions during February. No decisive rulings were issued during the month.

5.3Mergers, acquisitions, exits

Epic Games announced on February 22 the acquisition of Meshcapade, an AI startup specialising in the creation and animation of hyper-realistic digital human models from video recordings. The transaction extended Epic's investment in tooling for AI-generated character animation in real-time game and film engines.

Anthropic’s acquisition activity continued during the quarter with the publicly disclosed acquisition of Vercept, a software development startup founded in 2024. The transaction added to Anthropic’s 2025 acquisitions of Humanloop and Bun. OpenAI continued its Q1 acquisition programme.

Z.ai (Zhipu AI) and MiniMax progressed toward Hong Kong Stock Exchange listings during the quarter. Both companies reached public-market entry valuations above $6 billion, with formal listings completing during Q1.

5.4Infrastructure

Compute capacity remained the binding constraint on AI deployment. NVIDIA, AMD, and Intel announced expanded production schedules during the month, alongside continued progress on Google’s TPU partnership with Broadcom and Anthropic’s multi-year compute commitments.

Anthropic’s Series G announcement explicitly cited infrastructure expansion as a primary use of proceeds. CFO Krishna Rao’s statement framed the funding as enabling continued enterprise-grade product development at the scale demanded by Fortune-class customers. The pattern was consistent across the major frontier labs: capital deployment in 2026 is driven not principally by training compute requirements but by inference capacity to serve high-volume agentic workloads.

6Monthly Recap

Chronological summary of the month’s reportable events. Sources are linked in the corresponding section above.

7Outlook

OpenAI's release cadence implies a successor to GPT-5.3 in early-to-mid March. The naming may continue as GPT-5.4 or jump to a new generation. The model is expected to consolidate the GPT-5.3-Codex coding capabilities into the mainline rather than maintaining the separate Codex variant.

Mistral, having committed publicly to a $400 million ARR run-rate at a $13.8 billion valuation, is expected to ship product through March. The cadence of recent open-weight releases (Mistral Large 3 in December 2025, Ministral 3 family in late 2025) suggests at least one more flagship Mistral release before the end of Q1.

Replit’s Agent line is expected to receive a meaningful upgrade during March. The pattern of competing autonomous-coding tools — Cursor background agents, Claude Code subagents, Devin’s Agent Compute Units — suggests parallel-task execution will become a baseline expected feature rather than a differentiator.

The Digital Omnibus on AI is expected to reach political agreement during March. The legal effect of any postponement of high-risk obligations will then depend on a formal vote during April or May. Companies operating in or selling into the European market should treat the 2 August 2026 deadline as binding until any postponement is enacted.

Q1 venture-deployment data will close during March, with the OpenAI $122 billion round expected to be confirmed final by quarter-end. Reported AI share of total venture deployment exceeded 80% in independent estimates.

No decisive rulings are expected in the principal generative-AI copyright cases during March. Initial summary-judgement rulings remain on track for Q2 2026.