GOOGLE / DEEPMIND · MODEL RELEASE TRACKER
Gemini 3.8 Flash
Gemini 3.8 Flash is described in the supplied excerpts as Google’s prior fast (‘Flash’) model intended for speedy, lower-cost execution. It is repeatedly presented as the predecessor to Gemini 4 Hydrogen and is reported to have scored 18.2% on Terminal-Bench 4.1 in the cited coverage.CURRENT SNAPSHOT4/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
Not established from the available sources.
RELEASE TIMELINE
Published, source-backed release events only.
Google releases Gemini 4 Argon for robotics control, simulation, and policy improvement
Google released Gemini 4 Argon, a general-purpose physical-reasoning model aimed at robot control, Real2Sim scene construction, multimodal labeling of robot video, and iterative policy improvement. Google’s demos show Argon guiding a Boston Dynamics Spot and robot arms in MuJoCo twins, reconstructing scenes from teleoperation video, and synthesizing domain-randomized MuJoCo episodes (reported to raise success from 56% to 75% on one setup); Google reports Argon scores 87% on the ERQA benchmark, and makes it available to Google AI Ultra subscribers and via the paid Gemini API as gemini-4-argon, while publishing only limited per-demo failure data and relying on heavy prompt scaffolding in examples.
Thinking Machines unveils Inkling 2 Small, a 594B-parameter mixture-of-experts model
Thinking Machines says it has released Inkling 2 Small, a mixture-of-experts base model with 594 billion parameters (17 billion active) that accepts text and images and is initially offered via the Tinker platform; the company intends to publish the weights under an Apache 2.0 license after external safety testing. According to Thinking Machines and Artificial Analysis metrics cited by the company, Inkling 2 Small scores 31 on the Artificial Analysis AI Index (an 18-point gain over its predecessor), shows large gains on agentic and coding benchmarks, produces more output tokens and costs more per task, and claims the strongest performance among Western open-weight models.
Nvidia open-sources Flora, a Qwen3.5-based robot video labeling engine
Nvidia's GEAR lab open-sourced Flora, a data-engine composed of three fine-tuned Qwen3.5 models (9B for segmentation, 27B for descriptions, and a 27B 'Flora-Rewarder' for scoring) that labels robot videos for action boundaries, step-level rewards, and task success. Nvidia says Flora is already used in training its GR00T models at scale on GB300 GPUs and estimates labeling costs of about $0.72/hour versus ~$88/hour with GPT-6 Astra, though the cost figures are projections and benchmark comparisons include methodological caveats.
MiniMax releases M3.1 Flash reasoning model; open weights planned for Oct. 23
MiniMax has released M3.1 Flash, a reasoning model that scores 42 on the Artificial Analysis Intelligence Index (up 13 points from M3), supports up to 1 million-token context, and is claimed to run three times faster with lower cost per task (~$0.22). The model shows improvements on complex-reasoning and factual benchmarks (48% on "Humanity's Last Exam," 56% on AA-Omniscience) but reportedly has a high hallucination rate (93%); API pricing is published and MiniMax plans an open-weights release on October 23.
Google DeepMind unveils Gemini 4 Argon, a frontier model with 1M-token context
Google DeepMind announced Gemini 4 Argon, a new frontier multimodal model optimized for long-horizon reasoning and complex workflows. Argon is rolling out initially to trusted cyber defenders via the Fairwind Program, expands context length to 1 million tokens, reports leading benchmark performance across coding, finance, legal, and video understanding, and will be made more widely available after phased safety testing and engagement with U.S. government pre-release processes; Google also published introductory pricing.
VERIFIABLE FACTS
Every value stays attached to a source and date.
Announced by Google in September 2026 as part of September AI updates (listed alongside other Gemini 3.8 variants).
Delivers substantial gains in advanced, multistep reasoning and agentic tasks (including software engineering), improving over 3.7 Flash by executing extra reasoning steps and iteratively calling tools to produce more accurate outputs.
Designed to deliver Flash-level latency (low-latency Flash performance).
Described as multimodal; used in visual projects (e.g., image transformations and multi-camera prototypes) and noted alongside new audio models released the same month.
In the paper's evaluations, Gemini 3.8 Flash scored higher in several tests than StreamVibe-VL and other open-source models.
Described in the article as proprietary to Google.
Reported to have a 90% cache discount on cached input tokens (90 percent cheaper cached inputs).
In a published evaluation, loading full profession-specific profiles with Gemini 3.8 Flash produced 1.5–2.3× as many output tokens and increased estimated cost per successful call by 2.2–4.5×. On 60 tool-using BioMysteryBench problems, mean solve rates were 46.7% with the profile vs. 56.7% at baseline (difference −10.0 percentage points), driven by more frequent token- and time-limit stops under the profile for the tested model.
Gemini 3.8 Flash was listed among six models described as having web and email access in the Vals AI experiment.
Gemini 3.8 Flash was one of six models tested by safety researcher Matilda Glynn-Henley and Vals AI in an experiment that tasked AI agents with securing a £350 train refund; agents in the experiment were allowed up to twelve simulated weeks and up to 40 million tokens.
Scored 41 points on the Artificial Analysis Intelligence Index in a comparison reported by The Decoder.
Scored 18% on the CritPt benchmark in the same comparison.
According to the source, on the public WGO-Bench Gemini 3.8 Flash had higher pure description accuracy than Nvidia's Flora (Flora 'falls behind' Gemini 3.8 Flash).
Gemini 3.8 Flash is used "under the hood" in Genie 4 to help with its real-time simulations, acting as a planning/steering component for the world model.
According to DeepMind reporting, Gemini 3.8 Flash "watches the scene continuously, breaks a goal into steps, and uses new frames to check whether each step worked; if the world changes or the user steps in, the model adjusts its plan."
Evaluated on StateBench, a benchmark for long‑context state tracking; the model was included among the frontier models assessed by StateBench.
StateBench reports that Gemini 3.8 Flash (among other frontier models) "breaks down" on harder forms of state tracking as inputs grow longer, indicating reduced performance on more complex, longer dependency chains.
Gemini 3.8 Flash was one of seven models evaluated in a study that had models act as both 'writer' and 'grader' to create, run, and iteratively improve complete data science projects (downloading, cleaning, analyzing data, and writing a short paper).
In the study, the increase in average score by grader across the three revision rounds was 13 percentage points for Google's Gemini 3.8 Flash.
The article reports that Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index and describes that result as "roughly on par with Google's Gemini 3.8 Flash," implying comparable performance on that index.
Reported as 'roughly on par' with Thinking Machines' Inkling 2 Small (which scores 31) on the Artificial Analysis Intelligence Index.
The article states Inkling 2 Small is 'still far cheaper than Gemini 3.8 Flash', implying Gemini 3.8 Flash is more expensive per task.
Described as Google’s previous fast / 'Flash' model and the predecessor to Gemini 4 Hydrogen; Hydrogen is presented as a standalone upgrade for teams migrating from Gemini 3.8 Flash.
Reported Terminal‑Bench 4.1 score of 18.2% for Gemini 3.8 Flash.
Reports state that Gemini 4 Hydrogen completes tasks 2.5× faster than Gemini 3.8 Flash, indicating that Flash was slower in task completion compared with Hydrogen.
Reported that Hydrogen reduced hallucinations by a factor of five compared with Gemini 3.8 Flash, implying Gemini 3.8 Flash had a higher hallucination rate in the cited comparison.
WHAT CHANGED