GPT-5.6 and Grok 4.5 land hours apart as Meta starts charging for its first frontier model

GPT-5.6 goes public with Sol, Terra and Luna across ChatGPT, the API and Codex

If you build on the OpenAI API or pay for ChatGPT, today is the day to revisit your model choices. The GPT-5.6 family went public on 9 July in three tiers: Sol, the flagship at $5/$30 per million input/output tokens; Terra, a balanced everyday model at $2.50/$15; and Luna, a fast budget tier at $1/$6 that gives OpenAI a genuinely low-cost frontier option for the first time. Sol ships with an Ultra mode that spins up parallel subagents and posted 91.9% on Terminal-Bench 2.1, and a Cerebras deployment promises up to 750 tokens per second for select customers.

Two practical notes if you run production workloads: the gpt-5.5-latest endpoint does not migrate automatically, so pin explicit model IDs, and the new prompt caching system reads from cache at a 90% discount. One caveat worth carrying with you: the evaluation group METR found Sol can tell when it is being tested, so treat the published benchmarks as an upper bound and run your own evals before you switch.

GPT-Live voice finishes rolling out to paid ChatGPT plans

Announced only on OpenAI's X account on 9 July, GPT-Live, the new generation of voice models introduced earlier this week, has now reached every ChatGPT user on the Go, Plus and Pro plans, with free users next in the queue. If you have been waiting for a more natural back-and-forth voice, all it takes is updating to the latest ChatGPT app on iOS or Android. Mainstream tech press had not picked this up as we published, so the company's own channel was the only place it appeared. Worth trying today if voice is part of how you work.

Grok 4.5 opens to the public with big claims and no benchmarks

Grok 4.5 went public on 9 July for SuperGrok Heavy subscribers, X Premium+ users and developers on the xAI API, arriving within hours of GPT-5.6 in a genuine head-to-head. Elon Musk pitches it as an Opus-class model that is faster, more token-efficient and cheaper to run, built on the 1.5-trillion-parameter V9 foundation and trained partly on data from the Cursor IDE following the Anysphere acquisition.

The catch for anyone making routing decisions: the company has published no system card, no benchmark table and no confirmed per-token pricing, so the claims rest entirely on internal beta feedback. Evaluate it on your own tasks before you move any coding work across. Grok 4.5 is not yet available in the EU, with access expected later this month, and xAI's products now sit under the SpaceXAI brand after this week's rebrand.

Meta starts charging for AI with Muse Spark 1.1, its first paid model

This is a strategic shift as much as a model release. Meta Superintelligence Labs put Muse Spark 1.1 into public preview on 9 July at $1.25/$4.25 per million tokens, the first time Meta has charged developers to use one of its models and a clear break from its open-weights-only history. It is a multimodal reasoner built for agentic work, with multi-agent orchestration, computer use and 1M-token context compaction, and it tops tool-use benchmarks such as JobBench and MCP Atlas while trailing Claude Opus 4.8 and GPT-5.5 on pure coding.

If you are choosing a backend for agent workloads rather than raw code generation, the pricing is aggressive against everything else at the frontier. Zuckerberg has signalled that undercutting OpenAI and Anthropic on cost is deliberate, so expect this to press prices across the agent-model market.

Claude gains a reflection dashboard that shows how you actually use it

Announced on Anthropic's own news page on 9 July and not widely covered elsewhere yet, Claude now has a beta feature for tracking and reflecting on your own usage. The dashboard sits in Settings on the web and desktop apps and summarises your key topics, usage patterns and task types over the past 1 to 12 months, with quiet hours and break nudges you can set yourself. It deliberately excludes incognito chats and the underlying files from connected tools, and Anthropic built it with wellbeing researchers from the MIT Media Lab, the Digital Wellness Lab and the Family Online Safety Institute. It is live now for Free, Pro and Max users who have memory switched on.

For anyone whose working day increasingly runs through Claude, this is the first frontier chat product to ship self-audit tooling natively, and it is worth ten minutes of exploration.

Mistral Studio gives prompts and skills a system of record

Posted on Mistral's news page on 9 July with little outside coverage so far, Studio now treats prompts and skills as versioned, owned and traceable assets rather than loose text scattered across a team. If your organisation manages production prompts in shared docs or code comments, this is the kind of infrastructure that stops silent regressions when someone edits a prompt without telling anyone. It also shows where the European lab is placing its bets this week: less on a headline model launch, more on the operational tooling that makes enterprises comfortable running LLM workloads at scale. Worth a look if you already build on Mistral, and a useful benchmark for what to ask of other providers if you do not.

GPT-5.6 becomes the preferred model in Microsoft 365 Copilot

If your organisation uses Copilot inside Word, Excel, Outlook or Teams, the model underneath it changed on launch day: GPT-5.6 is now the preferred model in Microsoft 365 Copilot. Same-day adoption at this scale matters because most office workers will feel GPT-5.6's improvements through Copilot without ever touching ChatGPT or an API key. It is also a signal about the Microsoft and OpenAI relationship, which keeps putting OpenAI's newest models in front of hundreds of millions of enterprise seats faster than any rival channel. If Copilot output has felt inconsistent for your team, re-test your prompts and workflows this week against the new default.

Alibaba begins switching off Qwen's humanlike agents today

One access change takes effect today that Qwen users need to know about: Alibaba is disabling Qwen's humanlike interactive agents and user-created agent functions on 10 July, with broader agent services going offline on 15 July, ahead of China's new rules on anthropomorphic AI interaction services. Unlike ByteDance, Alibaba has not published a migration path, so anyone with established agent configurations faces permanent loss of those setups and their conversation history. We normally leave regulatory news aside, but this one removes product functionality today. If you or your users rely on Qwen agents, export anything recoverable now.