AI Model Daily - 2026-07-31

30-Second Summary

• OpenAI officially cut API prices for GPT-5.6 Luna and GPT-5.6 Terra: Luna is down 80% to $0.20/$1.20 per 1M input/output tokens, Terra is down 20% to $2/$12, and GPT-5.6 Sol now has an API Fast mode that OpenAI says is up to 2.5x faster at 2x the Standard price.
• Arena launched AutoEval: a reward-model-based early evaluation layer trained on millions of real Arena preference votes, covering Text, Vision, Image, Code, and other tracks. Image Edit Arena also gained the same 7 fine-grained categories as Text-to-Image.
• Google DeepMind released Gemini Robotics 2 / Gemini Robotics ER 2, moving Gemini Robotics from tabletop tasks into humanoid whole-body control, five-finger dexterity, and multi-robot collaboration. ER 2 is available through the Gemini API and AI Studio.
• Thinking Machines released Inkling-Small, an open-weight 276B total / 12B active MoE model that supports text, image, and audio. Artificial Analysis gives it an Intelligence Index score of 40, close to Inkling's 41.
• Visual and video models continue to move fast: Artificial Analysis ranks NVIDIA Cosmos3-Super-Image2Video-4Step as the #1 open-weight Image-to-Video model; Luma Layers is live in Luma Agents on Uni-1; Gemini Spark is starting to integrate with Chrome auto browsing.

Must Watch

1. OpenAI GPT-5.6 price and speed update: Luna/Terra price cuts, Sol Fast mode

• Event type: Official release / X first signal / availability update.
• First / official / available time: OpenAI's official blog and @OpenAI posted the update on 2026-07-30. API pricing became effective on 2026-07-30, with AWS pricing rolling out afterward.
• Capability change: GPT-5.6 Luna is down 80% to $0.20 input / $1.20 output per 1M tokens. GPT-5.6 Terra is down 20% to $2 input / $12 output per 1M tokens. GPT-5.6 Sol now has Fast mode, replacing Priority Processing; OpenAI says it is up to 2.5x faster with unchanged intelligence and costs 2x the Standard price.
• Availability: OpenAI API, Codex, and ChatGPT Work. OpenRouter also applied a 50% discount on top of the Luna/Terra cuts, showing Luna at $0.10/$0.60 and Terra at $1/$6 per 1M input/output tokens.
• Impact: This is not a new model launch, but it directly changes the cost curve for agents, code review, batch processing, and high-frequency classification workloads. Luna now moves into the pricing zone where strong small and mid-sized models compete. Internal benchmarks should rerun GPT-5.6 Luna/Terra on code execution, tool use, long-context work, and automated review.
• Sources: OpenAI blog, GPT-5.6 overview, OpenAI X, OpenAIDevs X, OpenRouter discounts, OpenRouter X.

2. OpenAI ARC-AGI-3 postmortem: two API settings moved GPT-5.6 Sol from 13.3% to 38.3% on the public set

• Event type: Official release / benchmark postmortem / X first signal.
• First / official / available time: OpenAI published the article on 2026-07-29, and it continued to circulate on X inside this window. @xiaohu posted a Chinese explainer on 2026-07-30 04:12 UTC.
• Capability change: OpenAI says GPT-5.6 Sol scored 13.3% RHAE on the ARC-AGI-3 public set under the official harness. With retained reasoning and compaction enabled, the score rose to 38.3%, while output tokens fell by 6x. OpenAI recommends API developers use the Responses API, preserve reasoning, and enable compaction.
• Availability: OpenAI Responses API. ARC-AGI-3 public tasks can be tried through ARC Prize.
• Impact: The point is not that the model suddenly became stronger in isolation. The key lesson is that eval harness design can dominate agentic benchmark outcomes. Future model comparisons must record retained reasoning, context compaction, rolling truncation, and prompt/harness details, or the same model can appear to have very different capability.
• Sources: OpenAI ARC-AGI-3 post, OpenAI X, xiaohu X, ARC Prize GPT-5.6 Sol results.

3. Arena AutoEval: model arena adds a reward-model early evaluation layer

• Event type: Arena appearance / official release / leaderboard methodology update.
• First / official / available time: @arena and the Arena official blog launched AutoEval on 2026-07-30. The Leaderboard Changelog also recorded the Image Edit Arena category update on the same day.
• Capability change: AutoEval trains pointwise reward models on Arena's large corpus of real user preference votes, then converts candidate model responses into soft votes and Arena Scores. Arena says that when two models are separated by at least 10 Arena points, AutoEval strongly agrees with the final live leaderboard ordering. The method extends to text, vision, image generation, code, and modality-specific reward models.
• Availability: Arena Text leaderboard now shows 1 AutoEval. Single Image Edit and Multi-Image Edit now have the same 7 categories as Text-to-Image.
• Impact: AutoEval gives anonymous models and new checkpoints an earlier ranking signal before live votes reach stable sample sizes. It also introduces a new calibration question: how well the reward model represents current real user preferences. Future reports should distinguish live human vote results from AutoEval results.
• Sources: Arena AutoEval blog, Arena X, Leaderboard changelog, Arena Text leaderboard, Arena Image Edit leaderboard.

4. Gemini Robotics 2 / ER 2: Google pushes Gemini Robotics into whole-body humanoids and multi-robot collaboration

• Event type: Official release / X first signal / availability update.
• First / official / available time: Google DeepMind and Google Blog published the release on 2026-07-30. @GoogleDeepMind posted demos the same day.
• Capability change: Gemini Robotics 2 expands into humanoid whole-body motions, controlling Apptronik Apollo 2 for walking, bending, and object handling. It also supports the SharpaWave five-finger hand for tasks such as tying knots and sealing ziplock bags. Gemini Robotics ER 2 focuses on video understanding, task planning, progress tracking, moment finding, multi-robot collaboration, and low-latency bidirectional streaming.
• Availability: Gemini Robotics ER 2 is available through the Gemini API and Google AI Studio, and in private preview in Gemini Enterprise Agent Platform. Gemini Robotics On-Device 2 is available to trusted testers.
• Impact: This is the most important physical AI model release in this window. For robotics developers, the ER 2 developer access is more immediately practical than the humanoid demos: it exposes high-level video understanding, progress judgment, and orchestration as callable model capabilities, with likely early impact in simulation, remote operation, warehousing, and indoor service robotics.
• Sources: Google DeepMind blog, Google Blog - ER 2, On-Device 2 model card, Google DeepMind X, Apptronik X.

5. Thinking Machines Inkling-Small: 276B/12B active open-weight model released

• Event type: Official release / open-weight release / third-party benchmark verification / X first signal.
• First / official / available time: Thinking Machines released Inkling-Small on 2026-07-30. Artificial Analysis published independent evaluation the same day.
• Capability change: Thinking Machines describes Inkling-Small as an MoE Transformer with 276B total parameters and 12B active parameters. It supports native reasoning over text, image, and audio, variable thinking effort, and up to 1M context. Official evals report SWEBench Verified 80.2%, Terminal Bench 2.1 64.7%, and AA Index v4.1 40.0. Artificial Analysis independently gives Inkling Small an Intelligence Index score of 40, close to Inkling's 41, and identifies it as Apache 2.0 open weights with 276B total / 12B active parameters.
• Availability: Tinker Playground, Tinker fine-tuning, and Hugging Face weights. vLLM and SGLang community posts show day-zero support signals.
• Impact: This is another open-weight agentic/multimodal model that deserves testing after Kimi K3. Watch real inference memory, third-party serving availability, context-length differences by provider, audio/vision input quality, and whether the 12B active MoE design gives a useful cost edge in real agent loops.
• Sources: Thinking Machines blog, Artificial Analysis article, Thinking Machines X, vLLM X, LMSYS/SGLang X.

6. Visual and video creation models: Cosmos3-Super-4Step, Luma Layers, Gemini Spark/Chrome

• Event type: Arena appearance / official release / availability update / X first signal.
• First / official / available time: Artificial Analysis posted the Cosmos3-Super-4Step ranking signal on 2026-07-30. Luma launched Layers on 2026-07-30. Gemini App posted Gemini Spark + Chrome integration on the same day.
• Capability change: Artificial Analysis says NVIDIA Cosmos3-Super-4Step distills are the #1 open-weight Image-to-Video models and are strong in Text-to-Image open weights. The public leaderboard shows Cosmos3-Super-Image2Video-4Step leading open-weight Image-to-Video without audio at Elo 1267, and Cosmos3-Super-Text2Image (agentic) leading open-weight Text-to-Image at Elo 1219. The Hugging Face model card says the Image2Video 4Step model uses fixed 4-step sampling, reducing the recommended 50-step + CFG setup from roughly 100 model evaluations to 4. Luma Layers lets Luma Agents extract layers from any image, then edit the product, background, or headline while preserving the rest. Gemini Spark is starting to integrate with Chrome auto browse for U.S. Google AI Pro/Ultra users.
• Availability: Artificial Analysis Image/Video leaderboards, Hugging Face NVIDIA Cosmos3-Super-Image2Video-4Step, Luma Agents, and Gemini Spark + Chrome.
• Impact: Video-model competition is moving from single-generation quality toward inference efficiency, editability, agent workflows, and browser execution. Cosmos3's 4-step distill has clear throughput value. Luma Layers and Gemini Spark/Chrome are more product capability than base-model launches, but they affect real creative-agent workflows.
• Sources: Artificial Analysis X, AA Image-to-Video leaderboard, AA Text-to-Image leaderboard, NVIDIA HF model card, Luma Layers, Luma X, Google Blog - Gemini Spark Chrome, GeminiApp X.

Arena Watch

• Arena / AutoEval: Starting 2026-07-30, Arena began using reward models trained on millions of live votes as an early evaluation signal. The Text leaderboard now shows 1 AutoEval; future reports should explicitly separate AutoEval from live human leaderboard results. Sources: AutoEval blog, Text leaderboard.
• Arena / Image Edit: The Leaderboard Changelog records that Single Image Edit and Multi-Image Edit now have the same 7 categories as Text-to-Image. This improves comparability for image-editing models, but it is not a new model release. Sources: Leaderboard changelog, Image Edit leaderboard.
• Artificial Analysis / Inkling-Small: AA gives Inkling Small an Intelligence Index score of 40, one point below Inkling, and records it as a 276B total / 12B active open-weight reasoning model with text/image/speech input. Source: AA Inkling Small.
• Artificial Analysis / Agnes 2.5 Pro Alpha: AA gives Agnes 2.5 Pro Alpha an Intelligence Index score of 39, pricing at $0.45/$0.90 per 1M input/output tokens, 1M context, text/image/video input, and first-party API availability. Coding Index 58.8 is its strongest low-cost signal. Sources: AA Agnes article, Agnes model page.
• Artificial Analysis / Cosmos3-Super: AA public leaderboards show Cosmos3-Super-Image2Video-4Step leading open-weight Image-to-Video without audio at Elo 1267. In open-weight Text-to-Image, Cosmos3-Super-Text2Image (agentic) leads at Elo 1219, while the 4Step version is at Elo 1181. Sources: AA Image-to-Video leaderboard, AA Text-to-Image leaderboard.
• OpenRouter availability/pricing: OpenRouter's 50% GPT-5.6 Luna/Terra discount stacks with OpenAI's price cuts, showing Luna at $0.10/$0.60 and Terra at $1/$6 per 1M input/output tokens. This is an availability/pricing signal, not a capability leaderboard. Sources: OpenRouter discounts, OpenRouter X.

X Watch

• @OpenAI / @OpenAIDevs: Posted the GPT-5.6 price-performance update. The developer account emphasized "more intelligence per dollar" in the GPT-5.6 API and said ChatGPT app and Codex CLI auto review are moving from GPT-5.4 to GPT-5.6 Luna; combined with the new Luna pricing, auto review is expected to become about 10x cheaper. Sources: OpenAI X, OpenAIDevs X 1, OpenAIDevs X 2.
• @arena: Posted the AutoEval thread and the 7-category Image Edit Arena update. Credibility is high because both can be cross-checked against Arena's blog and changelog. Sources: Arena AutoEval X, Arena Image Edit X.
• @GoogleDeepMind: Posted the Gemini Robotics 2 thread, showing Apptronik Apollo 2 whole-body control, five-finger dexterity, and multi-robot collaboration. Sources: Google DeepMind X, multi-robot X.
• @GeminiApp: Posted Gemini Spark integration with Chrome auto browse, initially rolling out to U.S. Google AI Pro / Ultra subscribers. It also posted the macOS voice / Speak to Window update. Sources: Gemini Spark X, Gemini Spark rollout X, macOS voice X.
• @ArtificialAnlys: Posted three leaderboard signals: Cosmos3-Super-4Step, Inkling Small, and Agnes 2.5 Pro Alpha. Sources: Cosmos X, Inkling X, Agnes X.
• @thinkymachines: Confirmed Inkling and Inkling-Small are available on Tinker, Tinker Playground supports chat, and a limited-time discount is active. Source: Thinking Machines X.
• @OpenRouter: Carried OpenAI's price cuts and added a 50% exclusive discount for GPT-5.6 Luna/Terra. Source: OpenRouter X.
• @LumaLabsAI: Launched Layers in Luma Agents, powered by Uni-1. This is a product capability release, not a new foundation model launch. Source: Luma X.
• @FireworksAI_HQ: Posted and amplified depthfirst dfs-large1 security-model signals, saying Fireworks helped scale training and inference. It also announced a Kimi K3 on Fireworks webinar. Treat the security model as X plus partner signal until depthfirst publishes a model card or long-form official page. Sources: Fireworks dfs-large1 X, Fireworks Kimi K3 webinar.
• @xiaohu: Chinese auxiliary source covering OpenAI's ARC-AGI-3 settings postmortem, a Seedance 2.5 next-day release claim, LayerX visual-deception attacks, and an Altman interview. Seedance 2.5 is a "pending launch preview" until ByteDance/Seedance publishes an official page or model entry. Sources: xiaohu ARC, xiaohu Seedance 2.5, xiaohu Altman interview.
• @SakanaAILabs: Dream-Cubed continued to circulate inside the window, but the original launch was on 2026-07-29. Record it as continued observation, not a new 2026-07-31 model launch. Sources: Sakana Dream-Cubed X, Dream-Cubed blog.
• @AnthropicAI, @AIatMeta, @MistralAI, @Alibaba_Qwen, @Kimi_Moonshot, @deepseek_ai, @runwayml, @Kling_ai, @TencentHunyuan, @cohere: No new high-confidence foundation-model launch was visible in this window. Some account pages had X limitations or no new visible articles.

Pending Confirmation

• Seedance 2.5: @xiaohu says it will be officially released tomorrow, and Chinese AI circles already show Seedance 2.0/2.5 examples. No ByteDance/Seedance official release page or model entry was visible in this window, so this remains a pending launch preview.
• dfs-large1: Fireworks and depthfirst-related X signals describe it as a specialized cybersecurity model for vulnerability finding and validation, possibly post-trained from open weights. It still needs an official model card, benchmark, availability path, license, and confirmation of whether it can be called independently.
• Gemini Robotics 2: ER 2 has Gemini API / AI Studio access, but the actual availability of the whole-body VLA layer and On-Device 2 remains limited to trusted testers or partner hardware. Confirm whether developers can directly call the VLA layer.
• Inkling-Small: The official post says up to 1M context, while the Artificial Analysis article records 256K context. Confirm actual context limits across the model card, Hugging Face config, and serving providers.
• Cosmos3-Super-4Step: AA ranking and the Hugging Face model card give strong signals, but hosted-entry defaults still need confirmation: resolution, frames, commercial terms, real latency, and hardware requirements.
• Arena AutoEval: Future leaderboard items need to distinguish live votes from AutoEval. If AutoEval affects ranking, record sample size, reward-model disclosure, and whether it changes anonymous-model exposure timing.
• GPT-5.6 Luna/Terra: After the price cut, rerun internal coding-agent, review, batch extraction, and long-context RAG benchmarks. OpenRouter's discount is an availability promotion and should not be treated as OpenAI's long-term official price.
• Gemini Spark + Chrome: Current rollout is limited to U.S. Google AI Pro/Ultra users and is a product/agent capability rather than a base-model launch. Confirm Chrome auto-browse permissions, approval flow, failure recovery, and enterprise availability timing.