AI Releases & Pricing
▶ Play 🎧 Overcast Share by text

Claude Fable 5's Subscription Access Ends Today, and It Now Costs More Than Anything Anthropic Sells

Plus: GPT-5.6 is priced and ready but still stuck behind a government review, Tencent's Hy3 goes fully open, and why Sonnet 5's "same price" isn't the same bill.

Today is the day the fine print catches up with Anthropic's most turbulent model launch of the year. Here is what actually changed, what's still a rumor, and the current price of every model worth comparing.

Fable 5 leaves subscriptions today, at $10 / $50 per million tokens

Starting today, July 7, Claude Fable 5 is no longer included in Pro, Max, Team, or premium Enterprise plans. Every session now draws from a separate usage-credit balance priced at $10 per million input tokens and $50 per million output tokens, according to Anthropic's own pricing page. That's double Claude Opus 4.8's rate and the most expensive tier Anthropic has ever listed for a generally available model.

The context explains the price: Fable 5 has had one of the roughest six weeks of any model release this year. It launched June 9 as Anthropic's first Mythos-class model, got suspended globally within 90 minutes on June 12 after a Commerce Department export-control directive, stayed dark for 19 days, and only came back July 1 once the license requirement was withdrawn, per Anthropic's redeployment notice. The six-day free window that followed (July 1 to July 7) was Anthropic's way of buying time while it builds out capacity, including new compute from a SpaceX-hosted data center deal. That grace period ends today.

If you're mid-project: enable usage credits in the Claude Console and set a spending cap before your allowance runs out. There's no metered fallback. Once the balance hits zero, a running session simply stops, whether that's mid-planning-pass or mid-week. Prompt caching cuts the input side by 90% ($10 down to $1 per million tokens on stable context), which is the single biggest lever if you're staying on Fable 5 for its 1-million-token context and 128k max output. Everything else should probably route to Sonnet 5 or Opus 4.8.

GPT-5.6 pricing is public. Access to GPT-5.6 is not.

OpenAI previewed GPT-5.6 on June 26: three models called Sol (flagship), Terra (balanced), and Luna (fast, cheap), with pricing already set at $5.00 / $30.00 per million tokens for Sol, $2.50 / $15.00 for Terra, and $1.00 / $6.00 for Luna (all with roughly 10% cached-input discounts). The catch is that none of it is available to ordinary users. At the U.S. government's request, OpenAI limited the launch to about 20 pre-approved partner organizations, and its own Help Center is explicit: "We have not announced a general-availability date." GPT-5.6 is not in ChatGPT and has no public waitlist.

That has not stopped a wave of rumors that OpenAI will flip the switch this week, specifically July 7 through 9, timed to land the same day Fable 5's billing cliff hits. That timing claim traces back to unnamed leakers and a Codex code leak reported by outlets like Odaily and BigGo, not an OpenAI announcement. Treat it as a rumor: plausible given the competitive setup, but OpenAI has given no confirmed date as of this morning. Worth checking your API dashboard for new model IDs before you plan around it either way.

Tencent's Hy3 goes fully open, and free for its own agent feature

Tencent officially released Hunyuan Hy3 on July 6, upgrading a preview that had been running since April, according to Tencent's own announcement and Caixin Global's reporting. It's a 295-billion-parameter mixture-of-experts model (21B active) released under Apache 2.0, meaning full commercial use and self-hosting rights, on Hugging Face, ModelScope, and GitHub. Tencent priced the API at 1 yuan per million input tokens and 4 yuan per million output tokens (roughly $0.14 / $0.56), with a cache-hit input rate of 0.25 yuan, and rolled it into its consumer Yuanbao assistant for free. It's the same open-weights playbook DeepSeek and Zhipu's GLM have been running: undercut the closed labs on price to force the market to compete on cost, not just benchmarks.

Fine print: Sonnet 5's token prices didn't go up. Your bill still might.

This isn't new today, but it belongs in the pricing record because it changes what "$2 per million tokens" actually means. When Anthropic launched Claude Sonnet 5 on June 30, it shipped a new tokenizer that breaks the same English text into roughly 27% to 42% more tokens than Sonnet 4.6 used for identical content, independently measured by developer Simon Willison and confirmed in Anthropic's own benchmarking data reported by The Decoder. Combined with Sonnet 5's more agentic behavior (it runs roughly three times as many tool-use loops as its predecessor on agent benchmarks), Artificial Analysis measured the real cost of an average task at $2.29 on Sonnet 5 versus $1.97 on the pricier-per-token Opus 4.8. Anthropic did the same thing with Opus 4.7's tokenizer last year. If you're budgeting off the $2/$10 introductory rate (which itself expires August 31 and rises to $3/$15), measure your own workload's actual token count before you trust the sticker price.

Watch: Gemini 3.5 Pro targeting July 17, still unconfirmed by Google

Google has not published a release date, but multiple reports, including AIbase and the same Odaily sourcing above, point to July 17 as Google's internal target, after the model already slipped once from a June window Sundar Pichai promised at I/O. Leaked specs mention a 2-million-token context window, double what Sonnet 5, Opus 4.8, and Fable 5 currently offer, but neither the context size nor pricing has been confirmed on Google's side. File this under "worth watching," not "worth budgeting around" yet.

Current prices, per million tokens (verified today, July 7, 2026)