AI news: May 2026

AI news in May 2026 clustered around two moments: Google I/O, where Gemini 3.5 Flash became the default model across Google’s stack, and the last week of the month, when Anthropic shipped Claude Opus 4.8 with an effort selector plus Claude Code’s dynamic workflows, and Mistral relaunched Le Chat as Vibe. GPT-5.5 Instant, meanwhile, became the model every ChatGPT user sees.

Gemini 3.5 Flash, Google’s new default

Google introduced Gemini 3.5 Flash on May 19 at I/O, generally available immediately: free as the default model in the Gemini app and Search’s AI Mode, and in the API at $1.50 per million input tokens and $9 per million output. Google cites 76.2% on Terminal-Bench 2.1, 83.6% on MCP Atlas and output generation 4 times faster than other frontier models, with Shopify, Salesforce and Databricks as early partners. Gemini 3.5 Pro was announced “for next month”, but independent sources report a delay to mid-July over an architectural rebuild; the Gemini Spark personal agent stayed in beta for Ultra subscribers in the US.

Google’s effective flagship being a cheap, fast Flash rather than the delayed Pro says where the competition is heading: cost and speed for agents, not peak scores. The benchmarks are Google’s own.

Source: Google DeepMind · also on TestingCatalog and TechCrunch

Claude Opus 4.8 arrives with an effort selector

Anthropic launched Claude Opus 4.8 on May 28, 41 days after Opus 4.7, available from day one in Claude.ai, Claude Code and the API on all plans, with no price increase ($5 per million input and $25 per million output; fast mode, 2.5x quicker, drops to a third of its previous cost). The new interface element is an effort selector next to the model picker, with high, extra and max levels. Anthropic cites 92.4% on Terminal-Bench 2.1 and 82.3% on OSWorld-Verified, and says it is about 4 times less likely to overlook code failures than Opus 4.7; Cognition CEO Scott Wu confirms it fixes the previous version’s verbosity and tool-use issues.

The effort selector turns a prompting trick into a product decision: same model, cost and latency tuned per task. Teams running Claude in production will need to decide which effort level each kind of task gets.

Source: Anthropic · also on TestingCatalog

GPT-5.5 Instant: ChatGPT’s new default

OpenAI rolled out GPT-5.5 Instant on May 5 as ChatGPT’s default model for all users. Per OpenAI, it produces 52.5% fewer hallucinated claims than GPT-5.3 Instant on high-stakes prompts (medicine, law, finance) and answers about 30% shorter. It also debuts memory-backed personalization: search can consult past conversations, files and Gmail, showing which memory sources it used, initially web-only for Plus and Pro. The system card classifies it as the first Instant model with “high” capability in biology/chemistry and cybersecurity, and documents regressions on the internal sexual-content and gore benchmarks. In the API it is the new chat-latest; GPT-5.3 stays available for only three more months.

Two operational notes: any product on the chat-latest alias changed models on May 5, and the three-month window for GPT-5.3 puts a deadline on pending migrations.

Source: OpenAI · also on the system card and TechCrunch

Gemini Omni: video generated and edited by conversation

Also at I/O, Google announced Gemini Omni Flash, the first model in the Omni family: it takes text, image, audio and video combined in one prompt and generates or edits video while keeping characters and lighting consistent across successive conversational edits, with physics understanding for more realistic effects and an avatar feature using the user’s own voice and appearance. Everything carries a SynthID watermark. Available immediately to Google AI subscribers via the app and Flow, and free on YouTube Shorts; clips are capped at 10 seconds for now and API access was announced for the following weeks.

Conversational editing is the workflow change: iterating a video by talking instead of regenerating from scratch. With 10-second clips and no API in May, serious use was still on hold.

Source: Google DeepMind · also on TestingCatalog

Mistral Medium 3.5 and remote agents in Vibe

Mistral introduced Mistral Medium 3.5 on May 22: a dense 128B model with a 256k context, open weights under a modified MIT license and self-hostable on 4 GPUs, scoring 77.6% on SWE-Bench Verified per the vendor. The API costs $1.50 per million input tokens and $7.50 per million output. Remote agents also land in Vibe, its coding CLI: asynchronous cloud sessions launchable from the terminal or Le Chat, with a teleport function to move a local session to the cloud, plus Le Chat’s Work mode in preview, which handles multi-step tasks with explicit approval before sensitive actions.

An open model at this level, self-hostable on 4 GPUs, gives teams whose data cannot leave their infrastructure a European option to evaluate; the remote agents cover a feature so far offered only by the US coding CLIs.

Source: Mistral AI · also on MarkTechPost and InfoQ

Nano Banana 2 and Nano Banana Pro reach general availability

Google Cloud confirmed on May 29 the general availability of Nano Banana 2 (Gemini 3.1 Flash Image) and Nano Banana Pro (Gemini 3 Pro Image) at 1K and 2K resolutions; 4K remains in preview, as does the new video-to-image generation (it accepts a file or a public YouTube URL as context). The preview versions shut down on June 25. Partners cited include Adobe (Firefly Enterprise), WPP with clients like L’Oréal and Unilever, and Shopify. Access with an enterprise SLA runs through the Gemini Enterprise Agent Platform.

With general availability and an enterprise SLA, image pipelines on these models can now be committed to clients. The preview shutdown date forces endpoint migrations before June 25.

Source: Google Cloud · also on TestingCatalog

Dynamic workflows in Claude Code: hundreds of subagents per task

On May 28 Anthropic shipped Claude Code’s Dynamic Workflows, generally available on Pro, Max, Team and Enterprise plans and via the API, Bedrock, Vertex and Foundry. Claude drafts a plan, spreads the work across tens or hundreds of parallel subagents, cross-checks the results and iterates to consensus; the user confirms before execution. It activates on request or through the ultracode setting. Anthropic’s cited example: Bun’s migration from Zig to Rust, some 750,000 lines, completed in 11 days with 99.8% of tests passing. Both Anthropic and the community flag token consumption well above normal sessions.

For large codebases it changes the scale of what fits in a single task (migrations, audits, refactors). Anthropic’s own recommendation applies: start with scoped tasks and monitor token spend.

Source: Anthropic · also on TestingCatalog and InfoQ

Grok Build: xAI’s terminal coding agent

xAI launched Grok Build, its terminal coding agent, in early beta: on May 14 for SuperGrok Heavy subscribers ($299 per month) and on May 25 for all SuperGrok and X Premium+ users. The CLI is written in Rust, coordinates up to 8 parallel subagents, runs a plan-first loop with step-by-step approval and is MCP-compatible. On May 29 it published the grok-build-0.1 model on the API (public beta): a 256K context, over 100 tokens per second and $1 per million input tokens, $2 output. An independent review by Can Demir separates what it actually delivers from what is oversold.

In a month already crowded with coding agents (Opus 4.8, dynamic workflows, Vibe), xAI joins with aggressive pricing and a subagent architecture. It is in beta and tied to a Grok subscription, so for now it can only be evaluated inside that environment.

Source: xAI · also on Grok Build 0.1 on API and a Medium review

Le Chat becomes Vibe, Mistral’s agent platform

On May 28 Mistral relaunched Le Chat as Vibe, with two modes. Work: search across Google Workspace, Outlook, SharePoint, Slack and GitHub, data analysis with charts, documents via Canvas and reusable skills. Code: remote coding agents in a sandbox, integrated with GitHub for projects and pull requests, with sessions that survive closing the laptop, plus a free VS Code extension and a CLI with editable plans and per-session permissions. Pricing: free tier, Pro at €14.99 per month, Team at €24.99 per user.

It positions itself as a complete European alternative to the ChatGPT-plus-Codex bundle, priced in euros and wired to the suites companies already use. For teams with data-residency requirements it is now a direct alternative to evaluate.

Source: Mistral AI · also on InfoQ and The Decoder

Mistral enters physics simulation with Physics AI

On May 27 Mistral announced Physics AI, a new model class created after its acquisition of Austria’s Emmi AI (undisclosed amount), whose team of over 30 researchers joins Mistral. The foundation is the Large Engineering Models: foundation models trained on the laws of physics which, per Mistral, predict physical behavior (airflow, heat, material stress) from geometry and boundary conditions in seconds on a single GPU, compared to the hours or weeks traditional solvers need. Partners cited: ASML, Airbus, Safran and Siemens Energy. No launch date, pricing or beta status; contact form only.

For industry (aerospace, automotive, energy) the promise is iterating designs in seconds instead of waiting on simulations. With only a contact form available, there is nothing to evaluate yet beyond the named partners.

Source: Mistral AI · also on Mistral research and HPCwire