Anthropic-managed Claude Platform becomes generally available on Google Cloud

Google has open-sourced gemini-api, a single-binary Rust command-line tool (google-gemini/gemini-api-cli, Apache-2.0 licence) that wraps the Gemini API into seven composable shell commands for coding agents such as Claude Code, Gemini CLI, Cursor and Antigravity as well as human developers.

Mechanism and agent ergonomics

Unlike Model Context Protocol (MCP) servers that load every tool schema into the model’s prompt at session start, gemini-api uses progressive disclosure: an agent runs gemini-api --help to see the seven top-level verbs and inspects subcommand flags only when needed. Binary outputs such as generated images, video or audio are written to disk while only the file path is printed to stdout, keeping raw base64 payloads out of the agent’s context window. Every command supports --json, --dry-run, --output and stdin piping.

Supported capabilities

• generate: text and structured JSON via Gemini models, images via Nano Banana 2, video via Veo 3.1, speech across 30 voices and music tracks via Lyria 3.
• analyze, transcribe and embed: multimodal inspection of images, video, audio and PDFs (automatically routing files above 100MB through the Files API), timestamped speaker-diarized transcription and text or multimodal embeddings.
• agent, files and batch: autonomous multi-step web research via deep-research-pro on the Interactions API, remote file lifecycle management and asynchronous Batch API jobs at a 50% token discount.

Ideogram has launched Ideogram Vector (ideogram-vector-1), a native vector generation model available in the Ideogram web app and through two v2 API endpoints: POST /v2/svg/generate/ideogram-vector-1 for text-to-vector generation and POST /v2/svg/vectorize/ideogram-vector-1 for converting raster images into editable SVGs.

How it works

Rather than generating a bitmap and running a post-hoc contour tracer, Ideogram Vector produces structured SVG markup with grouped layers, clean anchor paths, smooth multi-stop gradients and crisp lettering. Users can upload style references or character and object reference images (style_reference_images and reference_images), select LOW, MEDIUM or HIGH quality presets, and immediately recolour palettes, isolate path groups or resize artwork infinitely in the browser or tools such as Figma and Illustrator.

Key numbers

• Text-to-vector API price: $0.13 per SVG at Low quality, $0.16 at Medium and $0.20 at High ($0.32 per SVG when using image references).
• Image-to-vector API price: $0.10 per vectorised image.
• Web app plans: Free tier included with daily limits; Plus is $15 per month billed annually ($20 monthly) and Pro is $42 per month billed annually ($60 monthly).

OpenAI has made the Decisions API (POST /v1/decisions) generally available on gpt-6-luna. Designed for workloads that need structured judgements rather than free-form prose, the endpoint accepts text, images or native JSON (input_json) and evaluates a batch of independent questions in parallel at up to 10 times the speed of gpt-6-luna through the Responses API.

Decision types and GA additions

Each question in the questions array specifies one of three typed schemas: predicate (returns a boolean outcome when a calibrated probability between 0 and 1 meets a configurable threshold), choice (selects one key from up to 20 options with per-option probability distributions) or score (assigns a 1-indexed integer across 2 to 10 ordered rubric levels, plus an expected value score_ev and level probabilities).

New for general availability are native GPT-Live triggers that fire mid-conversation events during voice sessions without blocking speech, prompt caching on shared input prefixes, image attachments inside per-question instructions, and enterprise compliance controls including Private Link, Zero Data Retention, HIPAA eligibility and US or EU regional processing.

Key numbers

• Batch capacity: up to 400 questions per request (doubled from 200 in public beta) and up to 50 images per request across a 128k context window.
• Input pricing on gpt-6-luna: $0.10 per 1M uncached input tokens and $0.01 per 1M cached input tokens (90% discount), with zero charge for output tokens.

At AGNTCon + MCPCon, Google Cloud introduced Agent Compute Engine, a managed sandbox service built on the open-source Agent Substrate (ate.dev, released for Google Kubernetes Engine on 15 September 2026) and Google’s custom Titanium Infrastructure Processing Units (IPUs).

MicroVM isolation and state persistence

Every agent session runs inside its own hardware-backed microVM managed by Cloud Hypervisor, providing a dedicated kernel boundary against prompt-injected shell commands or untrusted code execution. Titanium IPUs offload network and storage I/O from the host CPU, while copy-on-write memory sharing deduplicates base runtime pages across concurrent sandboxes and snapshots memory state so paused agents release compute while waiting on user input or external APIs, then restore into persistent Hyperdisk storage.

On GKE, the underlying Agent Substrate architecture delivers up to 10 times higher compute density and sub-500ms suspend-and-resume latencies at 500 activations per second across 200k agents.

Governance and security controls

Google Cloud also detailed governance updates across Gemini for business: an Agent Gateway and Agent Registry with cryptographic SPIFFE Agent Identity, OpenTelemetry traces in Cloud Monitoring, Envoy-based cross-cloud policy enforcement, local endpoint sandboxes for Claude Code and Antigravity CLI, and token budget rules that downgrade routine tasks to lower-cost models. Agent Compute Engine enters preview in the coming months; Google Cloud has not announced pricing.

At AGNTCon + MCPCon in Seattle, Anthropic announced that it has donated the Agent Skills open specification (agentskills.io) to the Agentic AI Foundation (AAIF), a directed fund of the Linux Foundation that Anthropic co-founded in December 2025 when it contributed the Model Context Protocol (MCP).

How Agent Skills and AAIF fit together

First released by Anthropic in October 2025 and published as an open specification in December 2025, an Agent Skill is a directory containing a SKILL.md file with YAML frontmatter (name and description), markdown procedural instructions and optional helper scripts or reference templates. Agents load only the short frontmatter at startup and read the full instructions on demand. Since December 2025, more than 70 platforms including ChatGPT, Cursor, GitHub Copilot, Gemini CLI, Goose, Snowflake and VS Code have adopted the format, with over 100k skills published on GitHub.

Anthropic will co-maintain the specification with community contributors as AAIF standardises enterprise skill discovery, versioning and distribution. In the same announcement, AAIF accepted three additional open-source projects: OpenAI’s Open Responses (openresponses.org), Hugging Face and Meta’s OpenEnv for reinforcement-learning environments, and Alibaba’s OpenSandbox, bringing AAIF to seven hosted projects and more than 195 member organisations.

Anthropic’s Frontier Red Team and verification startup Theorem (theorem.dev) have published a large-scale study showing that Claude can autonomously patch memory-safety bugs and write machine-checked formal proofs in the Rocq (formerly Coq) theorem prover for unmodified C, C++ and Rust libraries drawn from Google’s OSS-Fuzz benchmark.

Binary-level Rocq proofs

Rather than asking engineers to rewrite production code in a verification-specific language, the team compiled each OSS-Fuzz target into a standalone RISC-V Linux executable with AddressSanitizer enabled (median size 18,000 instructions) and loaded its machine code into Rocq alongside a formal RISC-V instruction-set specification. Claude agents working inside Claude Code fixed underlying bugs in the source code and then constructed inductive invariants proving that the compiled binary never reaches an invalid instruction, failed assertion or AddressSanitizer memory error on any byte sequence up to 64MB.

Across 239 entry points in 123 libraries, Claude produced roughly 180 million lines of Rocq (~600k lines per proof), published at theorem-labs/oss-verify-proofs.

Functional correctness milestones

Beyond crash-freedom, Claude verified functional correctness on two self-contained systems: it proved in 20 days (~$20k in tokens) that the Linux kernel’s x86 zlib_inflate implementation conforms to RFC 1950 and returns Z_DATA_ERROR on invalid streams, and it wrote a 19,000-line Rust Python interpreter verified in Verus that matched CPython on 99.95% of 1.7M random test programs.

Microsoft AI has published peer-reviewed studies of the Microsoft AI Diagnostic Orchestrator (MAI-DxO) in Science (doi:10.1126/science.aeh7930) and BMJ Digital Health and AI, alongside the open-source Sequential Diagnosis Benchmark (microsoft.github.io/sdbench).

Multi-agent debate and cost control

Unlike static medical benchmarks such as USMLE where all patient findings are handed to a model at once, SDBench forces diagnostic systems to start from an initial complaint, ask follow-up history questions and order laboratory or imaging tests sequentially while tracking cumulative US healthcare test costs.

MAI-DxO wraps frontier language models in a structured virtual panel of five specialist roles: hypothesis generation, test selection, devil’s-advocate challenge, cost stewardship and quality checklist review. Across rounds of chain-of-debate reasoning, the panel updates a differential diagnosis and weighs expected information gain against the dollar cost of each test before committing to a final diagnosis.

Clinical evaluation next steps

On 304 New England Journal of Medicine (NEJM) clinicopathological cases, MAI-DxO achieved 83% diagnostic accuracy compared with 22% for 21 practising US and UK physicians working without reference materials (3.7 times as accurate) and 36% for physicians with textbooks and internet search (2.3 times as accurate); a cost-constrained configuration reached 81% at less than half the test spend. Microsoft AI is now launching a prospective clinical study with Beth Israel Deaconess Medical Center in Boston.