Skip to main content
Original: Swyx · 13/02/2026

​
Summary

The article discusses significant advancements in AI models, including Gemini 3 Deep Think, Anthropic’s funding success, and OpenAI’s GPT-5.3-Codex Spark, highlighting their performance metrics and implications.

​
Key Insights

“Gemini 3 Deep Think reaches new SOTA levels while also being very efficient - 82% cheaper per task.” — Discussing the efficiency and performance of Gemini 3 Deep Think.
“OpenAI rolled out their answer to Claude’s fast mode with GPT-5.3-Codex-Spark, delivering >1000 tok/s.” — Highlighting the performance of OpenAI’s new model.
“MiniMax M2.5 claims an Opus-matching 80.2% on SWE-Bench Verified.” — Referring to the performance claims of MiniMax M2.5.

​
Topics

​
Full Article

China open model week kept going with MiniMax M2.5 claiming an Opus-matching 80.2% on SWE-Bench Verified, however, as often happens on Thursdays, all 3 leading US labs had updates - Anthropic closed their 380Broundconfirmingahistoric>10xingofrevenueto380B round confirming a historic >10xing of revenue to 14B as of today (remember in August Dario projected 10B), with Claude Codes ARR doubling, hitting 2.5B year to date. Not to be outdone, OpenAI rolled out their answer to Claudes fast mode (2.5x speedup) with GPT-5.3-Codex-Spark, which delivers >1000 tok/s (10x speedup), an impressively fast turnaround of the Cerebras deal. All fantastic news, but we give the title story to the new Gemini 3 Deep Think today, and Jeff Dean dropped by the studio to give an update on the general state of GDM:This is the same model that scored that IMO Gold last summer, and is simultaneously the #8 best Codeforces programmer in the world and helping new semiconductor research, but perhaps most impressive is that it reaches new SOTA levels (eg on ARC-AGI-2) while also being very efficient - 82% cheaper per task - something Jeff was very excited about in his pod.AI Twitter RecapGoogle DeepMinds Gemini 3 Deep Think V2: benchmark jump + science/engineering reasoning mode shipping to usersDeep Think V2 rollout + access paths: Google is shipping an upgraded Gemini 3 Deep Think reasoning mode to Google AI Ultra subscribers in the Gemini app, and opening a Vertex AI / Gemini API early access program for select researchers/enterprises (GoogleDeepMind, Google, GeminiApp, tulseedoshi). Multiple Googlers emphasized this is meant to be a productized test-time compute heavy mode rather than a lab-only demo (OriolVinyalsML, JeffDean, demishassabis, sundarpichai).Key reported numbers (and whats notable about them):ARC-AGI-2: 84.6% (promoted as new SOTA; independently certified/verified by the ARC community) (Google, arcprize, fchollet, scaling01).Humanitys Last Exam (HLE): 48.4% without tools (sundarpichai, _philschmid, JeffDean).Codeforces Elo: 3455 (framed as only ~7 humans above it; discussion about no tools conditions and what that implies for evaluation) (scaling01, YouJiacheng, DeryaTR_).Olympiad-level written performance in Physics/Chemistry (and references to IMO/ICPC history) (Google, NoamShazeer, demishassabis, _philschmid).Cost disclosures for ARC: ARC Prize posted semi-private eval pricing like 13.62/task for ARC-AGI-2 and 7.17/taskforARC−AGI−1(arcprize).Real−worldengineeringdemosandclaimedimpact:SeveralpostspushthemessagethatDeepThinksvalueisinpracticalscientific/engineeringworkflows:findingerrorsinmathpapers,modelingphysicalsystemsincode,optimizingsemiconductorcrystalgrowth,andevenasketchCAD/STLpipelinefor3Dprinting(e.g.,laptopstandandturbine−blade−esquecomponents)(Google,Google,Google,GeminiApp,joshwoodward,tulseedoshi,OriolVinyalsML).ARCcontext/whatsaturatingARCmeans:FranoisChollet(ARCscreator)bothcelebratedcertificationandlaterreiteratedthatARCspurposeistosteerresearchtowardtest−timeadaptation/fluidintelligence,nottoproveAGI(fchollet,fchollet).InaseparatethreadhedefinesAGIastheendofthehumanAIgapandarguesbenchmarksmustevolveuntilhumanscannolongerproposetaskswheretheyoutperformAI,witharoughexpectationof 2030forthatstate(fchollet,fchollet).Opencoding/agentmodelsshippingfast:MiniMaxM2.5+ZhipusGLM−5battleforbestopenagenticcoderMiniMaxM2.5:distribution+positioning:MiniMaxsnewmodelispushedasanagent−verse/long−horizonagentmodel,rapidlyappearingacrossaggregatorsandtools:OpenRouter(OpenRouterAI),Arena(arena),IDE/agentslikeCline(cline),Ollamacloudfreepromo(ollama),Eigentagentscaffolds(EigentAI),Qoder(qoderaiide),andBlackboxAI(blackboxai).Benchmarkscitedinthethreadincludeclaimslike80.27.17/task for ARC-AGI-1 (arcprize).Real-world engineering demos and claimed impact: Several posts push the message that Deep Thinks value is in practical scientific/engineering workflows: finding errors in math papers, modeling physical systems in code, optimizing semiconductor crystal growth, and even a sketch CAD/STL pipeline for 3D printing (e.g., laptop stand and turbine-blade-esque components) (Google, Google, Google, GeminiApp, joshwoodward, tulseedoshi, OriolVinyalsML).ARC context / what saturating ARC means: Franois Chollet (ARCs creator) both celebrated certification and later reiterated that ARCs purpose is to steer research toward test-time adaptation / fluid intelligence, not to prove AGI (fchollet, fchollet). In a separate thread he defines AGI as the end of the humanAI gap and argues benchmarks must evolve until humans can no longer propose tasks where they outperform AI, with a rough expectation of ~2030 for that state (fchollet, fchollet).Open coding/agent models shipping fast: MiniMax M2.5 + Zhipus GLM-5 battle for best open agentic coderMiniMax M2.5: distribution + positioning: MiniMaxs new model is pushed as an agent-verse / long-horizon agent model, rapidly appearing across aggregators and tools: OpenRouter (OpenRouterAI), Arena (arena), IDE/agents like Cline (cline), Ollama cloud free promo (ollama), Eigent agent scaffolds (Eigent_AI), Qoder (qoder_ai_ide), and Blackbox AI (blackboxai).Benchmarks cited in the thread include claims like 80.2% SWE-Bench Verified and strong performance vs closed models in coding settings; multiple tweets stress throughput + cost as differentiators (e.g., 100 tokens/s and 0.06/M blended with caching are cited by Cline) (cline, cline, guohao_li, shydev69). Community vibe checks (e.g., Neubig) claim its one of the first open-ish coding models hed seriously consider switching to for daily work (gneubig).GLM-5: model scale + infra hints + open model leaderboards:Tooling ecosystem reports: GLM-5 is used on YouWare with a 200K context window for web projects (YouWareAI); one user reports ~14 tps on OpenRouter (scaling01).A more detailed (but still third-party) technical summary claims GLM-5 is 744B params with ~40B active, trained on 28.5T tokens, integrates DeepSeek Sparse Attention, and uses Slime asynchronous RL infra to increase post-training iteration speed (cline). Another tweet nitpicks terminology confusion around attention components (eliebakouch).Local inference datapoint: awnihannun reports running GLM-5 via mlx-lm on a 512GB M3 Ultra, generating a small game at ~15.4 tok/s using ~419GB memory (awnihannun).Arena signal: the Arena account says GLM-5 is #1 open model in Code Arena (tied with Kimi) and overall #6, still ~100+ points behind Claude Opus 4.6 on agentic webdev tasks (arena).A long Chinese-language-style analysis reposted via ZhihuFrontier argues GLM-5 improves hallucination control and programming fundamentals but is more verbose/overthinks, suggesting compute constraints (concurrency limits) show through (ZhihuFrontier).OpenAIs GPT-5.3-Codex-Spark: ultra-low-latency coding via Cerebras (and why UX becomes the bottleneck)Product announcement: OpenAI released GPT-5.3-Codex-Spark as a research preview for ChatGPT Pro users in the Codex app/CLI/IDE extension (OpenAI, OpenAIDevs). Its explicitly framed as the first milestone in a partnership with Cerebras (also touted by Cerebras) (cerebras).Performance envelope:The headline is 1000+ tokens per second and near-instant interaction (OpenAIDevs, sama, kevinweil, gdb).Initial capability details: text-only, 128k context, with plans for larger/longer/multimodal as infra capacity expands (OpenAIDevs).Anecdotal reviews highlight a new bottleneck: humans cant read/validate/steer as fast as the model can produce code, implying tooling/UX must evolve (better diffs, task decomposition, guardrails, agent inboxes, etc.) (danshipper, skirano).Model size speculation: There are community attempts to back-calculate size from throughput vs other MoEs; one estimate suggests ~30B active and perhaps 300B700B total parameters (scaling01). Treat this as informed speculation, not an official disclosure.Adoption/availability: Sam Altman later says Spark is rolling to Pro; OpenAI DevRel notes limited API early access for a small group (sama, OpenAIDevs). There are also Spark now with 100% of pro users type rollout notes with infra instability caveats (thsottiaux).Agent frameworks & infra: long-running agents, protocol standardization, and KV-cache as the new scaling wallA2A protocol as agent interoperability layer: Andrew Ng promoted a new DeepLearning.AI course on Agent2Agent (A2A), positioning it as a standard for discovery/communication across agent frameworks, mentioning IBMs ACP joining forces with A2A and integration patterns across Google ADK, LangGraph, MCP, and deployment via IBMs Agent Stack (AndrewYNg).Long-running agent harnesses are becoming product features:Cursor launched long-running agents and explicitly ties it to a new harness that can complete larger tasks (cursor_ai).LangChain folks discuss harness engineering research: forcing self-verification/iteration, automated context prefetch, and reflection over traces as levers that change outcomes materially (Vtrivedy10).Deepagents added bring-your-own sandboxes (Modal/Daytona/Runloop) for safe code execution contexts (sydneyrunkle).Serving bottlenecks: KV cache & disaggregation:PyTorch welcomed Mooncake into the ecosystem, describing it as targeting the memory wall in LLM serving with KVCache transfer/storage, enabling prefill/decode disaggregation, global cache reuse, elastic expert parallelism, and serving as a fault-tolerant distributed backend compatible with SGLang, vLLM, TensorRT-LLM (PyTorch).Moonshot/Kimi highlighted Mooncakes origins (Kimi + Tsinghua) and open-source trajectory (Kimi_Moonshot).A surprisingly common theme: files as queues: A viral thread describes a reliable distributed job queue using object storage + a queue.json (FIFO, at-least-once) as a minimalist primitive (turbopuffer). Another tweet claims Claude Code agent teams communicate by writing JSON files on disk, emphasizing no Redis required CLI ergonomics (peter6759).Research notes: small theorem provers + label-free vision training + RL algorithms for verifiable reasoningQED-Nano: 4B theorem proving with heavy test-time compute: A set of tweets introduces QED-Nano, a 4B natural-language theorem-proving model that matches larger systems on IMO-ProofBench and uses an agent scaffold scaling to >1M tokens per proof, with RL post-training rubrics as rewards. They promise open-source weights and training artifacts soon ( _lewtun, _lewtun, setlur_amrith, aviral_kumar2).LeJEPA: simplifying self-supervised vision: NYU Data Science highlights LeJEPA (Yann LeCun + collaborators) as a simpler label-free training method that drops many tricks but scales well and performs competitively on ImageNet (NYUDataScience).Recursive/agentic evaluation discourse: Multiple tweets debate recursive language models (RLMs) and stateful REPL loops as a way to manage long-horizon tasks outside the context window (lateinteraction, deepfates, lateinteraction).Top tweets (by engagement)Gemini 3 Deep Think upgrade + sketchSTL demo: @GeminiAppOpenAI Codex-Spark announcement: @OpenAI, @OpenAIDevs, @samaAnthropic funding/valuation: @AnthropicAIGemini Deep Think unprecedented 84.6% ARC-AGI-2: @sundarpichaiSimile launch + $100M raise; simulation framing: @joon_s_pk, @karpathy Read more

[AINews] Nano Banana 2 aka Gemini 3.1 Flash Image Preview: the new SOTA Imagegen model

Swyx · explanation · 80% similar

[AINews] GPT 5.4: SOTA Knowledge Work -and- Coding -and- CUA Model, OpenAI is so very back

Swyx · explanation · 79% similar

[AINews] Replit Agent 4: The Knowledge Work Agent

Swyx · explanation · 79% similar
Originally published at https://www.latent.space/p/ainews-new-gemini-3-deep-think-anthropic.