Grok 4.5: A Collaborative Model with a Price Edge

xAI and Cursor have co-trained Grok 4.5, positioning it as an Opus-class model that slots in between Opus 4.7 and 4.8 in capability. The headline here isn't raw performance — it's economics.

  • 6x cheaper than Anthropic's Opus models
  • 3x cheaper than GPT-5.5
  • Available directly in Cursor with increased usage limits
  • Accessible via API

The token efficiency makes it a compelling default for production workloads where cost-per-task matters more than marginal benchmark gains.

OpenAI Pushes GPT-5.6 and Reworks Voice

OpenAI is rolling out three GPT-5.6 variants — Sol, Terra, and Luna — to all users. Early feedback on Sol suggests it doesn't surpass Claude's Fable in raw intelligence, but its reliability is reportedly high enough that testers prefer it as a daily driver.

Separately, ChatGPT Voice is getting a genuine architecture overhaul with two new models: GPT-Live-1 and Live-1-mini. Key improvements:

  • The model can now respond while you're still speaking, eliminating awkward cutoffs
  • GPT-5.5 runs in the background, significantly boosting conversational intelligence
  • Overall smoother turn-taking compared to the old Advanced Voice Mode

Anthropic Extends Fable Access — With Caveats

Anthropic has extended access to Fable 5 through Claude subscriptions until July 12, capping usage at up to 50% of weekly limits. After that date, users will need prepaid usage credits to continue.

The extension is practical, but the shifting deadlines are eroding confidence in Anthropic's release commitments among some users.

Claude Cowork Goes Cross-Platform

A lighter version of Claude Cowork is coming to web and mobile. Sessions will run remotely and sync to your account, letting you start a task on a laptop and check in later from a phone. The full-featured desktop app experience remains unchanged. Beta rollout is expected over the coming weeks.

Meta Enters Multimedia Generation

Meta unveiled Muse Image and Video, a pair of new multimedia generation models. The image model shows strength in editing, text rendering, and web-grounded fact accuracy. A video generation model is currently in preview, with a broader release pending.

Other Notable Releases

  • Cognition released SWE-1.7, its strongest coding model yet, competitive with frontier models at reduced cost and running at 1,000 tokens/second
  • Cloudflare Drop lets you drag a folder into a browser and instantly get a live site — no account required
  • 64 Claude instances collaboratively rewrote the Bun JavaScript runtime in Rust in just 11 days
  • Perplexity is reportedly building an internal coding agent, tentatively called Perplexity Teammate, similar to Claude Code and OpenAI Codex