The 72-Hour Paywall
Grok 4.5 launched as a paid model on July 8. Three days later, xAI made it available as a free trial through the Grok Build CLI. Elon Musk posted the announcement on X, where it cleared 625,000 views within hours. No pricing document. No expanded terms of service. The binary: if you have the CLI, the model is free to try. (x.ai/news)
The access vector is specific and deliberate. The free trial runs through Grok Build, xAI's coding agent CLI, not through the chat interface and not through the API on paid terms. That surface choice is targeting a particular audience. Not general users who want a free alternative to ChatGPT. Engineers who want to run a real benchmark task on their own code before deciding whether Grok 4.5 belongs in their stack. The CLI is where that evaluation happens, and xAI is meeting those engineers there.
Grok Build v0.2.97 shipped the same day, and the changelog reads like infrastructure prep for the developer volume a free trial generates. Per-session token cost tracking now appears in headless JSON output, breaking down cost per prompt and per session, with Turn.usage exposed in the SDK. Voice mode unlocked for API-key sessions. Shell commands using ripgrep no longer require permission prompts by default, removing the most common friction point in automated script runs. Teams can now ship default allowed-command sets via managed_config.toml, with user deny-rules still winning. Rewind now fully removes the selected turn from both scrollback and the model's conversation history, not just the visible display. The update also fixed: IME text input attaching unrelated clipboard images on every character; queued prompts hanging on long blocking waits; crashes and session freezes. These are not cosmetic improvements. They are the class of changes you make before onboarding a developer population whose behavior you do not fully control. (x.ai/build/changelog, Releasebot)
The blast radius on the free trial: every developer who saw Grok 4.5 last week and declined to pay through X Premium. For context on the pricing landscape this week: OpenAI's GPT-5.6 Sol costs $30 per million output tokens. Anthropic's Fable 5, after the current free window closes, runs $50 per million output tokens. Grok 4.5 free trial via CLI is below Luna ($6 per million output tokens) in cost while positioned above Luna in claimed capability. The implicit argument is clear: evaluate before you lock a pricing relationship in.
xAI's release cycle is the real story. OpenAI held GPT-5.6 in government-restricted preview for weeks before the July 9 general launch. Anthropic has gated Fable 5 behind capacity constraints since late April. xAI went from paid to free developer trial in 72 hours. Grok 4.5 was co-trained on Cursor developer session data, which is a differentiation claim that requires actual developers to test. The 72-hour window is not altruism. It is a bet that if you benchmark Grok 4.5 on your own code before the evaluation period closes, you will benchmark it before you commit to Sol pricing.
managed_config.toml docs before setting up shared access.The Second Extension
Fable 5 was supposed to come off paid Claude plans on July 7. Anthropic extended it to July 12, hours before the cutoff. Today, hours before the July 12 cutoff, Anthropic extended it again, to July 19 at 11:59:59 PM PT. The extension applies to Pro, Max, Team, and enabled Enterprise seats: up to 50% of weekly subscription usage at Fable 5, drawing from the same pool as other Claude models. The Claude Code 50% usage limit boost runs through the same date. (Bleeping Computer)
The mechanism is supply, not promotion. Anthropic launched Fable 5 in late April and immediately ran into capacity constraints that prevented full-scale serving to the subscriber base. The free-window approach is the workaround: let paid subscribers evaluate Fable 5, build habits around it, then convert to usage credits once supply can support the load. Two consecutive last-minute extensions of one-week increments each means the supply curve has not yet reached the conversion point. The third estimate is July 19.
What this tells you: the demand signal from the evaluation window is strong enough that Anthropic will not cut the free lane until the paid lane can absorb the volume. The conversion price after July 19 is $10 per million input tokens and $50 per million output tokens. A lab does not extend a free promotion twice unless the alternative, watching users bounce off a $50 per million output price without a trial period, is worse for long-term retention than the cost of the extension.
The Rebuild
Google DeepMind scrapped the Gemini 2.5 Pro base model mid-sprint. The architectural rebuild, announced around July 7 to 8, targets a July 17 ship date, five days from today. The stated reasons: coding performance gaps, token efficiency failures, and long-task reasoning that fell short of Google's own standards for a flagship tier model. Enterprise testers reported benchmark performance that did not meet the competitive bar. The rebuild introduces a 2 million token context window, a Deep Think Reasoning Layer for complex problem-solving, and autonomous workflow capabilities. A ground-up rebuild at this scale costs hundreds of millions of dollars and forfeits months of development time. That is not a decision you make to gain two benchmark points. (BigGo Finance, HackerNoon)
The competitive context makes the timing legible. GPT-5.6 Sol landed on July 9 with strong coding benchmark scores. Anthropic's Fable 5 has been the capability anchor at the top of the frontier since late April. A Gemini 3.5 Pro that launched with the original 2.5 Pro base would have faced immediate benchmark comparisons it could not win. Google chose to delay rather than launch a flagship that loses on day one. That is a different call than most labs make in the race for launch-day press coverage, and it is worth crediting.
The departure context makes the week harder to read cleanly. Noam Shazeer, a senior architect at DeepMind, announced his departure for OpenAI on June 18. John Jumper announced his move to Anthropic on June 19. A ground-up architectural rebuild landed in the same four-week window as both departures. Google is not attributing the rebuild to the exits. The market is drawing its own inference, and that inference is showing up in Alphabet's valuation.
July 17 is Thursday. A full ground-up model rebuild shipping five days after a delay announcement would be an extraordinary pace. Watch whether the July 17 target represents GA launch or limited external testing.
The Velocity Map
After the most active 72-hour launch window of the summer, Sunday shows where each lab actually stands. OpenAI shipped the GPT-5.6 family (Sol at $5 per million input, Terra at $2.50, Luna at $1), GPT-Live voice, and ChatGPT Work on July 8 to 9. The pricing structure is set, the products are in market, and OpenAI has gone quiet. Meta shipped Muse Image, Muse Video with Content Seal watermarking, and Muse Spark 1.1 with the Meta Model API developer preview between July 7 and 9, then went silent. Mistral has not moved since Leanstral 1.5 on July 2.
xAI is the only lab that shipped product on July 11. Two items, both developer-facing. Anthropic shipped a policy update (the Fable 5 extension) but no product. Google is building. The map as of Sunday evening: OpenAI is consolidating around a three-tier pricing stack and watching adoption data. Meta launched its first image and video products and is tracking usage curves. xAI is in developer-acquisition mode, shortening its evaluation cycle to three days. Anthropic is buying time at the frontier tier while capacity builds. Google is making a costly bet that a rebuilt 3.5 Pro can compete with what the other labs already shipped last week.
Every lab on the frontier is managing a timeline they did not fully control. That is not a coincidence. That is what the frontier looks like at this pace.