Google updates Gemini 3 Deep Think, MiniMax releases M2.5

Google introduced an upgraded Gemini 3 Deep Think on February 12, 2026. The reasoning mode is designed for difficult problems in science, research, and engineering. It is available in the Gemini app to Google AI Ultra subscribers. Select researchers, engineers, and enterprises can also access it through a Gemini API early-access program for the first time.¹

The release updates an existing mode rather than introducing a new model family. Google reports an 84.6% score on ARC-AGI-2 and highlights potential uses such as analyzing complex research data and modeling physical systems.¹² The benchmark figures come from Google’s own materials. Companies considering an integration should also note that API access is initially limited to selected participants.¹

MiniMax introduced M2.5 on February 12 as an open-source model for coding, search, and office tasks.¹² The company says the model can use tools for these tasks and was openly available at launch.¹² MiniMax reported an 80.2% score on SWE-Bench Verified.¹ On February 15, it also announced the availability of M2.5-HighSpeed.³

MiniMax positions M2.5 as the next step after M2.1 and says its training included reinforcement learning across hundreds of thousands of complex, real-world environments.¹ The published coding results come from tests on MiniMax's own infrastructure using Claude Code as a scaffold; the company changed the default system prompt and averaged four runs.¹ The scores therefore are not independent confirmation of performance. MiniMax describes uses for office work, but its most concrete published figures concern coding and search.¹

Anthropic announced on February 12, 2026, that it had closed a $30 billion Series G funding round.¹ The company put its post-money valuation at $380 billion and said GIC and Coatue led the round.¹ Anthropic also reported a $14 billion annualized revenue run rate.¹

The valuation has more than doubled since its previous funding round in September 2025.² The deal shows how much capital investors are willing to commit to AI companies reporting rapid revenue growth.¹³ Anthropic identifies Claude Code as a major growth driver.¹³ A run rate is a projection based on current revenue, not revenue earned over a full year. Whether that growth will persist remains unclear. The published figures are company-reported, rather than drawn from independently audited financial statements.¹³

Claude Code accounted for about 4% of public GitHub commits in February 2026, according to an estimate by SemiAnalysis. The figure is not an official Anthropic metric and applies specifically to public commits, not all software changes. It therefore captures one visible part of the coding tool's use rather than measuring its full adoption.¹²

Anthropic said on February 12 that Claude Code had reached $2.5 billion in annualized revenue and that business subscriptions had quadrupled since the start of the year.³ Revenue and the estimated commit share describe different things: sales on one hand and visible GitHub activity on the other. The commit share alone does not show how much generated code reaches production or how much time companies save.

Spotify co-CEO Gustav Söderström said during the company’s fourth-quarter 2025 earnings presentation that some of its best developers had not written a line of code themselves since December.¹² Instead, they use AI tools to produce software changes while handling review and delivery.¹ His statement describes how those developers work, not a company-wide rule.¹²

Spotify uses an internal system called Honk, built on Anthropic’s Claude Code and integrated with Slack.¹² Developers can initiate tasks and review the results there.¹³ The example illustrates a shift in their work from writing code to directing and evaluating it.¹ It remains unclear how many developers work this way or what share of production code is AI-generated.¹²

EssilorLuxottica said it sold more than 7 million AI smart glasses developed with Meta in 2025. The figure includes Ray-Ban and Oakley models, but the eyewear maker did not disclose sales by product line.¹² It is EssilorLuxottica’s sales figure, not a separate shipment figure reported by Meta.¹

Sales rose sharply from the previous two years: EssilorLuxottica sold 2 million of the glasses across 2023 and 2024 combined.¹³ New models and additional AI features expanded the lineup in 2025.¹³ The increase points to a growing consumer market for AI-enabled wearables. The sales total does not show how often buyers use the AI features or how many units each product line contributed.¹²

IBM plans to triple entry-level hiring in the US in 2026.¹ Chief human resources officer Nickle LaMoreaux described the plan at a New York conference in February.² The roles are set to change: Rather than focus mainly on routine tasks, new hires would spend more time with customers, exercise human judgment, and review AI output.³² LaMoreaux pointed to junior software developers doing less basic coding and more customer-facing work.²

The plan contrasts with concerns that generative AI could reduce demand for early-career workers.³⁴ In HR, entry-level staff would step in when chatbots fall short, correct responses, and coordinate with managers.² It remains unclear how quickly the planned expansion will turn into hires; IBM has not disclosed an absolute hiring target.¹³

xAI reorganized its development work into four areas at a February 11, 2026, all-hands meeting: Grok, Coding, Imagine, and Macrohard.¹² Grok covers the chatbot and voice features, while Coding handles code generation and related infrastructure. Imagine focuses on image and video generation.¹² Macrohard is intended to develop AI agents that can use computers and automate work resembling that of a software company.¹²

The reorganization follows xAI’s merger into SpaceX and the departure of several co-founders.¹ It assigns distinct responsibilities to teams working on chatbots, media generation, and agentic software.¹² The meeting also touched on long-term ideas for computing infrastructure in space. No concrete expansion plan with confirmed capacity and timing emerged from those remarks.² For now, the team structure offers a clearer indication of xAI’s priorities than its infrastructure ideas.

Simile announced a $100 million Series A on February 12, 2026; a legal notice dated February 19 confirmed the round. Index Ventures led it, with Bain Capital Ventures, Hanabi Capital, and A* also participating.¹² The company is developing an AI platform intended to simulate human behavior and predict decisions. Reported potential uses include forecasting purchases and questions on corporate earnings calls.¹

Simile positions its simulations as a supplement or alternative to focus groups and traditional market research.³ CVS Health is testing the technology for decisions about store inventory and displays, according to reporting.³ The approach builds on Stanford-affiliated research into “Generative Agents.”³ How reliable the predictions are remains difficult to assess without publicly available performance metrics.¹³