Claude Sonnet 5 at $2 and $10, Fable 5 returns, Nano Banana 2 Lite

The week turned on one model launch on June 30 and one reversal the next morning.

Sonnet 5

Anthropic released Claude Sonnet 5 on June 30. It is the default model for Free and Pro and is also available on Max, Team and Enterprise, with the API ID claude-sonnet-5. Introductory pricing is $2 per million input tokens and $10 per million output, against $5 and $25 for Opus 4.8. The updated tokenizer can produce roughly 1.0 to 1.35 times more tokens for the same input. Cyber safeguards are on by default, and Anthropic reports that Sonnet 5 never produced a full working Firefox exploit in its testing, though partial successes ran slightly above Sonnet 4.6.

The post leans on charts for BrowseComp and OSWorld-Verified across effort levels instead of quoting numbers in text, so I can say less about raw capability than I would like.

Fable 5 comes back

Anthropic's release notes record access to Claude Fable 5 and Claude Mythos 5 restored on July 1, after the pause the launch page described on June 12.

Google's video and image pair

On June 30 Google released Gemini Omni Flash in public preview, a model for video generation and conversational editing that takes text, image and video, at $0.10 per second of output and capped at 10-second clips for now. Nano Banana 2 Lite, its fastest and cheapest image model, costs $0.034 per 1K image and produces text-to-image in about four seconds. Both carry SynthID watermarking.

Smaller items

Managed Agents picked up text deltas in event streams, per-session overrides of agent configuration, and webhooks for agent and deployment lifecycle events. Mistral released Leanstral 1.5 on July 2, an open 119B mixture-of-experts model (6B active) under Apache 2.0 for proof engineering in Lean 4. It reports 100% on miniF2F and 587 of 672 PutnamBench problems at its stated token budget. On the research side, Orca led the weekly papers list, and one paper argues a 35B agent can match trillion-parameter systems by scaling the horizon instead of the weights. That is a claim, not a result I have seen reproduced.

What I'd watch next: OpenAI's last flagship was GPT-5.5 in April, so any move there would reset the price picture.