Claude Fable 5 Came Back, Sonnet 5 Shipped, and the Bill Went Up Anyway

Three weeks ago I watched the best AI model on the planet get switched off by the U.S. government. Not deprecated. Not price-hiked into irrelevance. Switched off. A Commerce Department letter landed at Anthropic on a Friday afternoon and Claude Fable 5 went dark worldwide, including for Anthropic’s own staff. I’d been using it, I got the “this model is unavailable” wall, and I spent the back half of June rerouting everything to Opus 4.8 like it was 2025.

On July 1 it came back. And in the same 48-hour window, Anthropic shipped a brand-new model, Claude Sonnet 5, that looks on the pricing page like the deal of the summer. Two bucks per million input tokens. I almost repointed my whole Claude Code setup at it before my coffee finished brewing.

Then I did the math. You know how that goes. The frontier is back online, folks, but this week it comes with a warning label: bring a calculator.

• The Frontier Came Back Online
• Sonnet 5: Cheaper Per Token, More Expensive Per Answer
• Who’s Actually Shipping (And Who’s Stuck at the DMV)
• The Cheapskate Table (The Part That Didn’t Flinch)
• Horror Stories: The Meter Is Running
• What’s Coming
• The Takeaway

The Frontier Came Back Online

Let’s start with the good news, because there isn’t much of it in a month where the top model spent nearly three weeks as a paperweight.

Claude Fable 5, the one that’s been sitting at #1 on the LMArena text leaderboard all quarter, is available again as of July 1. The U.S. Commerce Department lifted the export controls it slapped on the model (and its unfiltered sibling Mythos 5) back on June 12. The whole mess started when Amazon researchers found a jailbreak: prompt Fable 5 the right way and it would happily identify a software vulnerability and, in at least one case, write the code to exploit it. That’s the kind of thing that gets you a government directive instead of a bug bounty.

The fix that ended the 19-day standoff is almost anticlimactic. Anthropic redeployed Fable 5 with a single new safety classifier trained to catch that one jailbreak technique. When a request trips it (reportedly 99%+ of the time), the request doesn’t just get refused. It gets quietly rerouted to Opus 4.8 for a second opinion. Commerce’s shiny new Center for AI Standards and Innovation reviewed the safeguard before pulling the controls.

And Fable 5 immediately went right back to running the table. It’s #1 on Arena Overall (1509), and #1 in Coding (1563), Creative Writing (1499), Instruction Following (1520), and Hard Prompts (1533). It’s #2 in Math by a single Elo point. Anthropic holds nine of the top twelve coding slots. Whatever you think about a model getting court-ordered off the internet, the thing is genuinely, stupidly good.

Here’s the catch nobody’s putting on the marketing slides: that reroute-to-Opus classifier is catching legitimate work. If you run security tooling, the honest kind, pentests you’re paid to run, vuln triage on your own codebase, your agent loop can hit that filter mid-run and get silently handed off to a different, weaker model with no clean signal that it happened. You asked Fable 5 to do a job and Opus 4.8 quietly answered. That’s a fun debugging session waiting to happen.

Sonnet 5: Cheaper Per Token, More Expensive Per Answer

Now the part where I almost got fooled.

Claude Sonnet 5 dropped June 30. Anthropic is calling it the most agentic Sonnet yet, and for once the marketing isn’t lying. It plans, it drives browsers and terminals, it runs autonomously at a level that a few months ago needed an Opus-sized model and an Opus-sized bill. It’s the default now for Free and Pro plans, it’s got a 1M-token context window, and on Artificial Analysis’s Intelligence Index it scores 53, good for #5 overall, sitting right on top of GPT-5.5 with high reasoning and just ahead of Opus 4.8 on agentic knowledge work.

The pricing looks like a gift: $2/$10 per million tokens as an intro rate through August 31, moving to $3/$15 after. Opus 4.8 is $5/$25. So Sonnet 5 is half the price of Opus, nearly as smart, and more agentic. Switch everything, right?

No. This is the whole point of the post. Look at cost per task instead of cost per token, and the deal evaporates.

Artificial Analysis actually ran the numbers, and an average task on their Intelligence Index costs:

• ~$1.20 on Sonnet 4.6 (the old one)
• ~$1.97 on Opus 4.8 (the “expensive” one)
• ~$2.29 on Sonnet 5 (the “cheap” one)

Read that again. The new Sonnet costs about 15% more per completed task than Opus 4.8, and roughly double the old Sonnet. The per-token sticker went down and the actual bill went up.

How? Two things stack. First, a new tokenizer that’s less efficient on a lot of content, so the same input can map to 1.0 to 1.35x more tokens than it did under the old scheme, depending on what you feed it. Second, the more agentic behavior itself. With max effort it burned about 40% more output tokens per task than Sonnet 4.6 and took roughly 3x the agentic turns on knowledge-work evals. A more agentic model does more work per task by design: longer chains, more tool calls, more self-correction, more retries when a step fails. All of it billed.

This is Anthropic’s move now, and the-decoder called it out by name: raise the real cost while the per-token rate on the pricing page stays flat or drops. It’s not a scam. Sonnet 5 is a legitimately strong model and for a lot of agentic work it’s the right call. But if you switched based on the headline number without measuring your own workload, you’d be paying more and thinking you saved. The pricing page lies. Measure cost per finished task on your representative work, or don’t complain when the invoice shows up.

Who’s Actually Shipping (And Who’s Stuck at the DMV)

Step back and the shape of the week is Anthropic having an absurdly good run while everybody else’s flagship is vaporware.

OpenAI’s GPT-5.6 (the Sol/Terra/Luna trio) previewed June 26, and then went nowhere you can reach. At the U.S. government’s request, OpenAI shipped it to about 20 vetted partner organizations first, gated behind a safety review, with general availability “coming weeks” (analysts are betting July 10–17). OpenAI publicly grumbled that this kind of gating “shouldn’t become the norm.” It’s a launch you literally cannot use.

Google’s Gemini 3.5 Pro announced at I/O on May 19 with a June target, and has now slipped again to mid-July. Some reports say a full architecture rebuild after early testers flagged token-efficiency and coding problems. Still preview-only.

And while Google’s flagship stalls, its people are leaving. In one week Anthropic picked off four-plus senior Gemini researchers, including Nobel laureate John Jumper of AlphaFold fame, plus a couple more on pretraining and coding tools. Noam Shazeer, one of the “Attention Is All You Need” authors, went to OpenAI. Alphabet shed something like $270 billion in market cap over the stretch, and Anthropic is now reportedly the most valuable private AI company on earth. The people who built Gemini are literally going to go build the next Claude.

So the scoreboard for the week: one lab restored the world’s best model, shipped a new one, and hired the competition’s brain trust. The other two are showing you a locked demo and a “give us until next month.”

The Cheapskate Table (The Part That Didn’t Flinch)

Here’s the thing about all this frontier drama. If you’re actually shipping software, none of it changed what you should be running day to day. The value layer barely moved. The Cheapskate method is the same as it’s been: take the leader’s Arena rating in each category, look at everything within about 50 rating points of it, and pick the cheapest one, because the top of every Arena category is compressed into a tiny band. You are almost never paying for a meaningfully smarter model at the top. You’re paying for a rounding error.

The headline of that table is GLM-5.2, Zhipu’s roughly 753B open-weight MoE, and it’s the model I’d actually point you at this week. It’s MIT-licensed, runs about $3 per million output tokens on open hosts, tops the open-weight SWE-bench Pro leaderboard at 62.1%, and it sits on Artificial Analysis’s Intelligence-vs-Cost Pareto frontier. Cheaper than Gemini 3.5 Flash while scoring higher on capability. When two totally different methods (my dumb “cheapest in the band” heuristic and AA’s Pareto math) point at the same model, that’s about as strong a signal as this game gives you.

One more nuance that rhymes with the Sonnet 5 lesson. Gemini 3.5 Flash wins Overall and Creative on per-token price, but per task it’s actually pricier than Gemini 3.1 Pro, because Flash burns a lot of reasoning tokens. AA has Flash costing about 75% more per task than 3.1 Pro despite the lower sticker. Same trap, different model. If your workloads are long and multi-step, price them on your own traffic before you trust the cheap-looking one.

And the throughline that’s held for a couple months now: US models have fallen from about 70% to about 30% of OpenRouter token volume in a year, with Chinese labs around 46% and six of the top ten by volume. When your #1 model can be switched off by a government letter, the open-weights model you can run yourself starts looking less like a budget compromise and more like insurance. Own your weights.

Horror Stories: The Meter Is Running

Two this week, and they’re the same story wearing different hats.

Agentic bills are blowing past budgets. TechTimes reported July 4 that Anthropic rushed enterprise spend controls out the door because customers’ agentic bills are overshooting what they planned. Seventy-eight percent of IT leaders reported unexpected charges from consumption-based AI pricing in 2026. Now drop a new, more-agentic default model into that fire. Sonnet 5 does more per task on purpose, and more-per-task means more-per-invoice. If you flip your team to it and don’t watch the meter, you find out at the end of the month.

Fable 5’s safety net is snagging real work. As covered up top, the reroute-to-Opus classifier is catching legitimate security workflows and silently swapping models on you mid-agent-loop. It’s a smaller-dollar horror than a runaway bill, but it’s the sneakier one, because nothing errors out. You just get a subtly worse answer from a model you didn’t ask for, and you’re left wondering why your pentest agent got dumber overnight.

Both are the same lesson the whole post keeps hammering. The number you look at is not the number you pay.

What’s Coming

• GPT-5.6 Sol/Terra/Luna hits general availability “coming weeks,” with analysts guessing July 10–17. OpenAI is promising Cerebras hosting at up to 750 tokens/sec, which, if it’s real, is a genuinely different speed tier for agent loops.
• Gemini 3.5 Pro slipped its GA to roughly July 17 after an apparent architecture rebuild. Believe it when you can actually call the API.
• DeepSeek V4 is rumored for a full public launch mid-July (V4 Pro at 1.6T total / 49B active, V4 Flash at 284B / 13B, 1M context). DeepSeek is already the single largest token author on OpenRouter, so a real GA matters.
• Portugal’s Amália shipped July 1, the first open LLM in European Portuguese, Apache 2.0 on Hugging Face, built for about €7 million in public money. The sovereign-model trend is quietly real, and “we built our own for the price of a mansion” is a hell of a flex.

The Takeaway

The frontier rebooted this week and it’s genuinely better than it was in June — Fable 5 is back, Sonnet 5 is a real step up in agentic muscle, and the lab that shipped both is running away with the talent war. If you only read headlines, this was a triumphant week for AI.

But the actual skill in 2026 isn’t picking the model at the top of the leaderboard. It’s reading the meter. Sonnet 5 taught it, Gemini 3.5 Flash taught it again, and the enterprise-spend-controls fire drill taught it to a lot of finance teams the hard way. Cost per token is marketing. Cost per finished task is reality. And they’re diverging on purpose.

So run GLM-5.2 for coding until something dislodges it. Price your cheap-looking models on your own traffic before you trust them. Keep an open-weights option in your back pocket for the week your favorite model gets a letter from Washington. And when a new model shows up half the price of the old one, smile, nod, and go check the invoice.

The leaderboard’s back. The meter never stopped.