Another busy day in artificial intelligence, headlined by a fresh batch of model releases and infrastructure moves. Google introduced Gemini 3.7 Flash, its fastest-iterating “workhorse” series yet; OpenAI and Cerebras unveiled an “Ultrafast” inference tier for GPT-5.6 Sol; DeepSeek opened up its agent harness to developers; Mistral shipped OCR 4.1; and Google outlined a new open-source compiler designed to make privacy-preserving AI practical. Here are the five stories shaping the week in AI.
Google unveils Gemini 3.7 Flash, a faster and cheaper workhorse for coding and agents
Google has introduced Gemini 3.7 Flash, which the company calls its “most intelligent workhorse model yet” for coding and agents. The release lands just three weeks after Gemini 3.6 Flash and reflects direct developer feedback, according to Tulsee Doshi, senior director of product management at Google, writing on behalf of the Gemini team.
The model posts substantial gains in software engineering, knowledge work, and web development. On the FrontierCode 1.1 Main benchmark, it scores 43.6% versus 3.6 Flash 34.4%; on DeepSWE v1.1 it reaches 65.3% versus 49.0%. It also improved on knowledge-dense workloads — 34.0% versus 22.0% on the GDP.pdf document-processing evaluation, and 30.4% versus 17.0% on AutomationBench for real-world business workflows. In web development it reaches an Elo of 1588 on Arena.ai WebDev Arena, up from 1538.
Notably, Gemini 3.7 Flash launches at an introductory price of half the original 3.6 Flash cost per million tokens, and Google demoed it orchestrating sub-agents for tasks ranging from generating real-time 3D game assets to training robotics models via multimodal agent loops.
OpenAI and Cerebras unveil “Ultrafast” mode for GPT-5.6 Sol
Cerebras and OpenAI have shared an early look at Ultrafast Mode, a new service tier launching first in the OpenAI API and powered by Cerebras hardware that delivers up to 750 output tokens per second without any quality compromise. The tier has opened to a select group of customers, with access expanding over time.
According to Cerebras benchmarks, GPT-5.6 Sol on Ultrafast runs roughly 11x faster than Anthropic Fable 5 and about 5x faster than Opus 4.8 on Fast mode. Cerebras put the model head-to-head on Humanity Last Exam, a 2,500-question benchmark: Sol Ultrafast worked through all questions in 11 hours and 11 minutes, while Claude Fable 5 took 78 hours and 27 minutes at comparable accuracy, a nearly 7x speedup. On GDP-Val, a benchmark for economically valuable knowledge work, Ultrafast delivered a 5.6x end-to-end speedup with no quality degradation.
DeepSeek launches an open-source agent harness
DeepSeek has released DeepSeek Harness in developer preview, an open-source agent harness built on the Cordis plugin system. Its tagline — everything is a plugin — reflects a design where every capability (models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI) can be swapped or recomposed via configuration without modifying source code.
The project emphasizes traceability: everything the model sees is recorded in an append-only session log, including system prompts, reasoning, tool calls, and subagent scheduling, which can be inspected via a Trajectory view. It ships with multiple runtime modes — Standard, Code mode, Minimal, and Creator mode for building custom agent presets. Developers can get started with: npx @deepseek-ai/dsh web.
Mistral ships OCR 4.1 with paragraph-level structure
Mistral AI has released OCR 4.1, the latest version of its document-understanding service. The release adds native paragraph-level bounding-box extraction, structural block labels, and block-level confidence scores — valuable for parsing mixed documents like invoices, contracts, and research papers.
OCR 4.1 is available as a public preview, with pricing at $4 per 1,000 pages and $5 per 1,000 annotated pages. It is accessible through Mistral dedicated OCR API endpoints including structured annotations and batched processing.
Google open-sources HEIR compiler to make private AI practical
Google has added HEIR (Homomorphic Encryption Intermediate Representation) to its Private Computing Toolkit — an open-source compiler designed to make cryptographically-secure private AI inference more practical. With homomorphic encryption, computations run directly on encrypted data, so a cloud service could deliver recommendations or analytics without ever seeing the underlying data.
Staff Software Engineer Jeremy Kun described the technology as rapidly maturing and shifting the privacy/capability trade-off from a security question to a question of cost — a cost that is rapidly decreasing. HEIR automates much of the work that previously required a team of cryptographers.
The bottom line
This week news underscores two converging themes across AI. The first is speed: OpenAI and Cerebras are racing on inference throughput, while Google continues its rapid Flash-series cadence pairing better benchmark scores with falling prices. The second is trust and control: DeepSeek pushes open traceable agent harnesses, Google invests in cryptographic privacy, and builders get cheaper faster ways to deploy frontier models.
