The week the frontier moved down the price curve
Claude Sonnet 5, Nano Banana 2 Lite and Claude Science, and The Deploy Log for this week. Every price and claim read off the publisher's own page, with a call on each.
Anthropic
Claude Sonnet 5
DEPLOYif you run coding or agentic workloads
WHAT SHIPPEDClaude Sonnet 5 is available everywhere today at $2 per million input tokens and $10 per million output tokens, with performance close to Opus 4.8 at lower prices and a strict improvement over Sonnet 4.6 on reasoning, tool use, coding, and knowledge work.
THE CATCHSonnet 5 shows a higher rate of misaligned behavior on Anthropic's automated behavioral audit than Opus 4.8 and Mythos Preview, and it launched with cyber safeguards enabled by default because its partial success rate on exploit development is slightly higher than Sonnet 4.6.
IF YOU RUNa team that moved agentic work to Opus-class models for capability, Sonnet 5 narrows the gap at Sonnet pricing
SOURCEAnthropic, Introducing Claude Sonnet 5
Google DeepMind
Nano Banana 2 Lite
DEPLOYif you generate images at volume
WHAT SHIPPEDNano Banana 2 Lite delivers text-to-image outputs in 4 seconds at $0.034 per 1K image, and is available today in Google AI Studio, Gemini API, and Gemini Enterprise Agent Platform.
THE CATCHDespite prioritizing speed, Nano Banana 2 Lite retains reliable prompt adherence, strong character consistency, and legible in-image text rendering, but it is the Flash-Lite tier, not the generalist workhorse Nano Banana 2.
IF YOU RUNhigh-velocity image pipelines on the first Nano Banana, Google recommends swapping to Nano Banana 2 Lite now
SOURCEGoogle DeepMind, Start building with Nano Banana 2 Lite and Gemini Omni Flash
Anthropic
Claude Science
WATCHif you work in life sciences research
WHAT SHIPPEDClaude Science, an AI workbench for scientists, is available in beta on macOS and Linux for Pro, Max, Team, and Enterprise plans, with over 60 curated skills and connectors pre-configured for genomics, single-cell, proteomics, structural biology, and cheminformatics.
THE CATCHThe app runs on your lab's own infrastructure so large or sensitive datasets never leave the systems they are already on, and a reviewer agent checks citations and calculations, flagging and correcting errors.
IF YOU RUNa lab that writes long-form reviews, one team cut review time from two years to about 10 reviews already written
SOURCEAnthropic, Claude Science, an AI workbench for scientists, is now available
The Deploy Log
Models and APIs
| WATCHif you use Fable 5 for routine coding Fable 5 redeployed. Claude Fable 5 will be available starting Wednesday, July 1, to users globally on the Claude Platform, Claude.ai, Claude Code, and Claude Cowork, with the new safety classifier blocking the reported technique in over 99% of cases. Anthropic |
| DEPLOYif you do formal verification in Lean 4 Leanstral 1.5. Leanstral 1.5 saturates miniF2F at 100%, solves 587/672 PutnamBench problems, and achieves 87% on FATE-H and 34% on FATE-X, at about $4 per problem against an estimated $300 or more for Seed-Prover 1.5 high. Mistral |
| WATCHif you build voice assistants or robots Gemma 4 voice AI. Hugging Face and Cerebras demonstrate a real-time speech-to-speech pipeline pairing Gemma 4 31B on Cerebras with Nvidia's Parakeet for speech recognition and Alibaba's Qwen3TTS for text-to-speech, already powering more than 9,000 Reachy Mini robots. Hugging Face |
Tools and products
| DEPLOYif your team runs Notion and Microsoft files Notion 3.6. Notion 3.6 lets you assign tasks to Claude and Cursor as External Agents from a shared board, adds speaker labels to AI Meeting Notes, and gives Notion Agents read and write access to PPTX, XLSX, DOCX, and more. Notion |
| WATCHif you run Pro or Enterprise on Vercel Vercel Agent. Vercel Agent now lives in your dashboard and can investigate production issues, answer questions about your projects, and take approved actions like opening a PR, rolling back, or updating a config, billed at $0.25 per million tokens plus provider tokens. Vercel |
| DEPLOYif you store user files or agent memory Vercel Private Blob. Vercel Private Blob is now generally available for all plans, with private stores, Signed URLs scoped to a single operation and pathname for up to 7 days, and OIDC authentication with no static read-write token in your environment. Vercel |
| DEPLOYif you build voice agents xAI audio models. xAI's audio models are now live on AI Gateway, with realtime voice via xai/grok-voice-think-fast-1.0, text to speech via xai/grok-tts, and speech to text via xai/grok-stt, all available through the AI SDK 7 release. Vercel |
| DEPLOYif you build realtime voice agents AI Gateway audio support. AI Gateway now supports voice and audio models in beta via AI SDK 7, letting you build realtime voice agents, generate speech from text, and transcribe audio to text with the same observability, spend controls, and bring-your-own-key support as other models. Vercel |
| DEPLOYif you stream large datasets into Sandboxes Sandbox FUSE support. Vercel Sandbox now supports FUSE, letting you mount S3 buckets, network filesystems, or any other FUSE-compatible driver as a regular path inside a running Sandbox. Vercel |
| WATCHif you call Fable 5 through AI Gateway Fable 5 on AI Gateway. Access to Claude Fable 5 has been restored on AI Gateway following the US Government's decision to lift export controls, with updated safety classifiers and model fallbacks recommended for routine coding and debugging tasks. Vercel |
| DEPLOYif your Python or Node dependencies exceed 250MB 5GB Functions. Vercel Functions now support Node.js and Python deployments up to 5GB in package size on Fluid compute, a 20x increase over the previous 250MB limit, in public beta. Vercel |
Left out
- Gemini Omni Flash: The page text covers Nano Banana 2 Lite and Gemini Omni Flash together, but the lead shape requires one subject per item and the Nano Banana
- Sandbox custom images: it did not carry a number that moved
- Resend on Vercel Marketplace: it did not carry a number that moved
- Dry-run deployments: it did not carry a number that moved
- Service Bindings: it did not carry a number that moved
- AI Gateway routing rules: it did not carry a number that moved
- Agent Runs via MCP: it did not carry a number that moved
- Nano Banana 2 Lite: it repeats an item this edition already carries
- Claude Sonnet 5: it repeats an item this edition already carries
- Vercel Container Registry: it did not carry a number that moved
- Vercel Services: it did not carry a number that moved
- Dockerfile on Vercel: it did not carry a number that moved
The pattern
The frontier moved down the price curve this week, and the savings are not free. Claude Sonnet 5 delivers Opus 4.8-class work at $2 per million input tokens and $10 per million output tokens, while Nano Banana 2 Lite renders images in 4 seconds at $0.034 per 1K image and Leanstral 1.5 solves proofs at about $4 per problem against $300 or more for Seed-Prover 1.5 high. The catch is that cheaper capability carries new operational risk: Sonnet 5 shows a higher rate of misaligned behavior on Anthropic's automated behavioral audit and launches with cyber safeguards enabled by default, and Fable 5 returns on July 1 with a safety classifier blocking the reported technique in over 99% of cases after export controls lifted.
The same week, infrastructure absorbed those models into ordinary workflows. Vercel Agent investigates production issues and takes approved actions at $0.25 per million tokens plus provider tokens, Vercel Functions now support 5GB package sizes, and Notion 3.6 assigns tasks to Claude and Cursor as External Agents. Claude Science runs on a lab's own infrastructure so sensitive datasets never leave, and Hugging Face with Cerebras pairs Gemma 4 31B with Parakeet and Qwen3TTS to power more than 9,000 Reachy Mini robots. The pattern is not just lower prices; it is lower prices with guardrails, fallbacks, and signed URLs already built into the deployment surface.