Gemini 4 Carbon tops the Artificial Analysis Intelligence Index v5 with a score of 59, initially opening access to select cyber defenders. Mistral Large 4 releases its trillion-parameter model weights, roughly matching GLM-5.3 on a coding agent benchmark while using 47% fewer tokens per task on average. Through multi-model orchestration, Copilot Fusion scores 2 points below Opus 5.5 paired with Claude Code at a cost of $5.59 per task, versus $11.42 for the comparison setup. In DecoyCount's test of 1,660 real-world images, all tested frontier models scored below 50% accuracy, and larger thinking budgets did not necessarily narrow the gap between prompted and unprompted performance.
Comparable benchmark scores can be pursued through lower token use or runtime orchestration.
Mistral has released improved weights for its trillion-parameter multimodal model, Large 4, with support for self-hosted deployment. Compared with the public preview released this month, the new version adds three weeks of reinforcement learning. Mistral says it scores 48% on the AA coding agent index, roughly matching GLM-5.3 while using 47% fewer tokens per task on average. The model also offers three reasoning modes so developers can balance performance and cost.
Read original →GitHub has introduced Fusion in Copilot, orchestrating Anthropic and OpenAI models at runtime. On the AA coding agent index, it scores 2 points below Opus 5.5 paired with Claude Code. Its cost per task is $5.59, versus $11.42 for the comparison setup, delivering a comparable benchmark result at about half the cost.
Read original →🔍 Analysis: Mistral Large 4 and Copilot Fusion take different routes to efficiency on the same AA coding agent index: the former optimizes token use within a single model, while the latter orchestrates Anthropic and OpenAI models, accepting a small score gap in exchange for a lower cost per task. For developers, model scores alone are no longer enough to determine the choice; actual task costs and deployment options also matter. Using 47% fewer tokens does not directly translate into a 47% reduction in total cost.
Gemini 4 Carbon sets a new overall ranking, while DecoyCount tests whether models can resist visual distractions.
Google has released Gemini 4 Carbon, which ranks first on the Artificial Analysis Intelligence Index v5 with 59 points. In internal testing, its success rate for secure code generation was more than twice that of Argon, released two weeks earlier. Access begins this week for select cyber defenders, followed by a gradual expansion to more trusted partners through the Fairwind Program.
Read original →The visual counting benchmark DecoyCount contains 1,660 real-world images and seven types of distraction mechanisms. It uses standard questions, questions with hints, and multi-image, multi-hop questions to distinguish perception errors from reasoning errors. All tested frontier models scored below 50% accuracy, with gaps of 11.8–22.1 percentage points between responses with and without hints. Stronger models, larger thinking budgets, and agentic workflows often widened this gap instead; adding scene context did not improve overall performance either.
Read original →🔍 Analysis: Gemini 4 Carbon's score of 59 comes from a broad intelligence index, while DecoyCount separates counting errors in real-world images into perception and reasoning failures. They do not measure the same kind of reliability. In DecoyCount, increasing the thinking budget often widens the accuracy gap between prompted and unprompted responses, suggesting that “thinking longer” is not a universal fix for visual tasks. Developers still need to validate performance against specific distraction mechanisms rather than use an overall ranking as a substitute for task-level acceptance testing.
H-Net v2 seeks to change tokenization, while MAX opens up its full stack for execution across accelerators.
Cartesia's H-Net v2 learns tokenization end to end from raw bytes and reports a 3–5× efficiency gain in controlled scaling experiments. In larger-scale experiments, it matched an open model trained on 11 trillion tokens using roughly one-tenth the data and compute. It can also learn to treat variable names as individual tokens and shows better scaling performance on code than BPE.
Read original →Modular has opened the full source code for MAX, including its compiler, runtime, and hardware backends. Developers can write code once and run it across different accelerators, with the release covering the entire execution stack. Commercial products built on MAX also require no licensing fees or royalties.
Read original →🔍 Analysis: H-Net v2 targets efficiency in how a model learns representations from bytes, while MAX puts control in the compiler, runtime, and hardware backends, addressing training and deployment respectively. The former's reported 3–5× efficiency gain cannot be directly extrapolated into lower production inference costs, and the latter's lack of licensing fees does not mean it is free to run. For teams building their own model stacks, training compute, hardware adaptation, and software licensing need to be accounted for separately.
OpenAI and ServiceNow expand two-way integration, while a Utah pilot provides a set of user-reported results.
OpenAI and ServiceNow are expanding their partnership to advance two-way integration around enterprise workflows. OpenAI's models and agents will be brought into ServiceNow. In the other direction, ServiceNow workflows will become accessible through ChatGPT and Codex, making the partnership more than simply connecting models to an enterprise platform.
Read original →Utah provided Claude Code to 34 developers across 10 state government agencies for a four-week pilot. Of the participants, 80% set up a working environment within 30 minutes, demonstrating how quickly they could get started. By the end of the pilot, 71% reported saving more than 20 hours in total; this time-saving measure was self-reported.
Read original →🔍 Analysis: OpenAI and ServiceNow are expanding enterprise integration from “platforms calling models” to “ChatGPT and Codex calling workflows,” while the Utah pilot tests whether tools deliver benefits once introduced into developers' jobs. These address two hurdles in adoption: whether business workflows can be connected, and whether those connections save time. Of the 34 developers, 71% reported saving more than 20 hours—a finding suitable for further verification, not a direct measure of organization-wide productivity gains.
Monthly merged pull requests reach 220 million, while monthly commits hit 7.38 billion.
GitHub released its Octoverse 2026 report at Universe: monthly merged pull requests tripled in less than a year to 220 million. Monthly commits more than tripled to 7.38 billion, with most of the increase concentrated in the second half of this year. New repositories created each month also doubled to 30.3 million. The report additionally introduces industry-level data and separately tracks repositories involving interoperability or reverse engineering.
Read original →