GPT-5.6 Finally Shipped, Then Grok and Meta Ate Its Lunch

Between Tuesday and Wednesday, July 8th and 9th, three different labs dropped flagship models. xAI shipped Grok 4.5. OpenAI finally let GPT-5.6 out of the building. Meta launched Muse Spark 1.1. Three frontier releases in about 48 hours, which is the kind of thing that used to happen once a quarter.

Nobody won. Not on intelligence, anyway. The smartest model in the world is still Claude Fable 5, same as last month. What every one of these launches was really fighting over was the bill. Not the sticker price on the API page. The actual amount of money that leaves your account when the agent finishes the job. That’s the whole story this week.

• The 48-Hour Flagship Pileup
• GPT-5.6 Shipped From House Arrest
• Grok 4.5 and the Number Nobody’s Talking About
• Meta Closed the Door on the Way Out
• The Cheapskate Picks
• Horror Stories From the Wild
• What’s Coming: July 17 Is Going to Be a Bloodbath
• The Takeaway

The 48-Hour Flagship Pileup

For the last two weeks I’ve been writing about GPT-5.6 as vaporware. OpenAI previewed it in late June and then locked it to about twenty government-vetted partners under a frontier-AI review process. You couldn’t use it. I couldn’t use it. It was a press release with a benchmark chart attached.

Then on July 9th it went fully generally available on ChatGPT, the API, Codex, and GitHub Copilot, all at once. Great. Except the day before that, xAI dropped Grok 4.5, a model priced so aggressively it makes GPT-5.6 Sol look like a luxury purchase. And the same day GPT-5.6 shipped, Meta launched Muse Spark 1.1 at roughly a quarter of what Anthropic and OpenAI charge for their flagships.

So OpenAI’s big moment, the flagship it had been dangling for a month, got sandwiched between two competitors on the same 48-hour window, both of them attacking the exact axis OpenAI is weakest on. If you’re keeping score at home on the Artificial Analysis Intelligence Index, the top now reads something like Fable 5 at 60, GPT-5.6 Sol at 59, Opus 4.8 at 56, Grok 4.5 at 54. That’s a dogpile inside six points. When the top four models are that close on capability, “which one is smartest” stops being an interesting question. “Which one is cheapest to actually run” becomes the only question.

GPT-5.6 Shipped From House Arrest

Let’s give OpenAI its due first. GPT-5.6 comes in three sizes and, unlike some past launches, the naming actually maps to something useful:

• Sol, the flagship: $5 in / $30 out per million tokens
• Terra, the everyday model: $2.50 / $15, which OpenAI pitches as “GPT-5.5 performance at half the price”
• Luna, the cheap one: $1 / $6

All three carry a 1M context window and 128k max output. Sol lands at #2 on the AA Intelligence Index at 59, and it set a new high on Agents’ Last Exam at 53.6, beating Fable 5 by a full 13 points. If you’d stopped reading OpenAI’s launch post there, you’d think they’d taken the crown.

Then you flip to SWE-Bench Pro, the coding benchmark, and Fable 5 scores 80% while Sol manages 64.6%. OpenAI’s response to that gap was, roughly, to question whether the benchmark is any good. Which is a bold move when it’s the same benchmark everyone was citing last month. Here’s the thing I keep coming back to: the headline eval and the embarrassing eval are on the same page of the same launch. That’s not a knock on the model, it’s a reminder that you pick the benchmark that matches your actual work, not the one the vendor put in bold.

Simon Willison had early access to Sol and his take was the most honest one I read: “definitely very competent,” but not more impressive than Fable for complex coding. He also noticed Sol likes to over-investigate before it acts, which is a polite way of saying it burns tokens thinking in circles before it does the thing. More on that theme in a second, because it’s the theme of the entire week.

One genuinely cool bit: OpenAI is putting Sol on Cerebras hardware at up to 750 tokens per second. Fast is a real feature for agent loops. Just not fast enough to be the story.

Grok 4.5 and the Number Nobody’s Talking About

Here’s where I have to admit I underestimated xAI.

Grok 4.5 launched July 8th at $2 in / $6 out per million tokens. That’s over 60% cheaper than Opus 4.8 or GPT-5.5. It lands at #4 on the AA Intelligence Index, above every open-weight model and above every Gemini model on the board too. It was trained on real Cursor session data, which is the kind of thing that either makes a coding model great or gets you into a benchmark scandal, and this week it did both. We’ll get there.

But the number nobody’s putting in their headline is cost-per-task. Artificial Analysis measured GPT-5.6 Sol at $1.04 to run one Intelligence Index task. Grok 4.5? $0.31. Roughly a third of the cost for a model that’s four spots down and change on raw intelligence.

How? Token efficiency. Grok 4.5 uses about 15,954 output tokens on a SWE-Bench Pro task. Opus 4.8 uses 67,020 for the same work. That’s a 4.2x difference in how much the model talks to itself to arrive at an answer, and since you pay per token, that ratio lands directly on your invoice. On coding-agent tasks specifically, AA clocked Grok at $2.49 per task versus $5.07 for GPT-5.5 and $11.80 for Fable 5.

Read that last sentence again. Fable 5 costs almost 5x what Grok costs to finish the same coding task. Not because its per-token price is 5x higher (it’s not), but because it thinks longer and harder and you pay for every one of those tokens. This is the trap I keep watching people fall into: they compare the $/million numbers on the pricing page, pick the “cheaper” model, and then get a bigger bill because it’s chattier. The sticker price is not the price.

Now the asterisks, because there are real ones:

Grok 4.5’s launch scores were vendor-reported with no model card shipped. Cursor had to withdraw one internal benchmark after it came out that Grok had trained on their codebase — which, given it was literally trained on Cursor session data, is exactly the conflict you’d worry about. And the capability tradeoff is ugly: on the AA-Omniscience test, Grok’s accuracy climbed from 35% to 52%, which sounds great, except its hallucination rate also jumped from 25% to 54%. It knows more, and it’s more confident when it’s wrong. That’s arguably the worse failure mode. A model that admits it doesn’t know is annoying; a model that fabricates with conviction is the one that ships a fake function name into your codebase at 2 AM.

Oh, and it dropped from a 1M context window down to 500k, and it’s not available in the EU yet. But at $0.31 a task, a lot of people are going to forgive all of that.

Meta Closed the Door on the Way Out

Muse Spark 1.1 is the quiet winner of the week and also the most ironic.

Meta priced it at $1.25 in / $4.25 out, about a quarter of flagship rates, threw in $20 of free credits per account, and it immediately started landing inside the competitive band across the Arena leaderboard. Overall #6 at 1490. Coding #8. Instruction Following #8. Hard Prompts #7. It’s not topping anything, but it’s competitive everywhere at a price that undercuts almost everyone. That’s the definition of a value pick.

Here’s the irony. This is Meta, the company whose entire AI identity was giving models away. Llama got downloaded over a billion times because you could run it yourself, fine-tune it, own it. Muse Spark 1.1 ships closed-weight. No download, no local deploy, no fine-tuning. If your reason for loving Meta’s models was that you could actually possess them, that reason is gone. The cheapest brand-name model on the board this week is also the one that just walked away from open source. Make of that what you will. I have some feelings about it, and none of them are inspirational.

The Cheapskate Picks

Okay, the part you can actually use. Same method as always: for each Arena category, I take the leader’s rating and find the cheapest model within about 50 rating points of it. Because Arena’s top end is compressed (the whole competitive pack usually fits inside a 50-point window), “cheapest in the band” is a real choice between models that are genuinely close, not settling for garbage.

One caveat this week: Grok 4.5 and GPT-5.6’s Terra and Luna tiers are so new that Arena hasn’t accumulated enough votes to rank them in most categories yet. So they don’t show up in the band even though they’re obvious value plays. I’ve noted where that matters. Opus-tier leaders are priced around $25/million output; Fable 5 leaders at $50.

A few notes. Muse Spark 1.1 is the story here. It wins three categories outright as the cheapskate pick, mostly because it’s brand-name cheap and Arena caught up to it fast. For coding, Qwen3.7 Max is the by-the-numbers pick, but the real coding value if you don’t care about the Arena ranking is Grok 4.5 at $2.49 a task. It’s just not Arena-rated yet, so I can’t put it in the table honestly.

And Gemini 3.5 Flash keeps quietly winning Math and Creative for the third or fourth week running. At $9 output it beats every Claude Opus-thinking variant on Math except one. If you’re doing math-heavy or creative work and you’re paying Fable 5’s $50, you are lighting money on fire for a rating difference smaller than the noise in the measurement.

Also worth flagging: Fable 5’s Math #1 rating of 1548 is sitting on only 360 votes. That’s thin. Don’t over-trust a fresh #1 with that little data behind it.

Horror Stories From the Wild

Every roundup needs the part where I tell you what’s going to ruin your day. This week it comes courtesy of Cursor, and it’s a good one.

Multiple users on the Cursor forums reported that when they selected Grok 4.5 in the CLI, their usage logs showed they were actually being billed for GPT-5.6-sol-medium. Separately, Grok 4.5 was routing its subagent calls to GPT-5.6-terra-medium instead of the model people thought they’d picked. So you select the cheap, token-efficient model, and you get silently billed for a different, pricier one. This is the exact nightmare scenario for a week where the whole selling point is cost-per-task. The tooling can’t reliably tell you which model you’re even running, let alone what it costs. There were also plain “Grok 4.5 errors out and won’t come back” reports. New model, new plumbing, same old chaos.

The second horror story is the Grok hallucination number I already mentioned: a 54% hallucination rate on AA-Omniscience, up from 25%. I’m repeating it here because it deserves to live in the horror section too. A model that’s wrong more than half the time on knowledge questions and more confident about it is not a model you point at anything that touches production without a human reading every line.

What’s Coming: July 17 Is Going to Be a Bloodbath

Mark the 17th. Two big things land the same day.

Gemini 3.5 Pro is finally targeting July 17, after Google scrapped its base model and rebuilt the whole thing from the ground up. Reportedly engineers found structural failures in recursive tool-calling and SVG generation and decided a rebuild beat a patch. Context window, pricing, specs: still unconfirmed. Google has reportedly shed something like $225B in market cap over the delays and a DeepMind talent exodus, so there’s a lot riding on this one not being a dud.

DeepSeek V4 graduates from preview to official stable release the same day. V4-Pro at 1.6 trillion params, V4-Flash at 284 billion, both with 1M context. And if you’re already using DeepSeek, note the deadline: the API migration is mandatory by July 24th. After that, deepseek-chat and deepseek-reasoner start returning errors. Don’t get caught by that on a Friday.

Also on deck: Grok 4.5 EU availability is promised for later this month, and Anthropic extended Fable 5 access on all paid plans plus kept Claude Code’s weekly limits 50% higher through July 19th.

Background hum, unchanged: DeepSeek is still around 17.6% of all OpenRouter tokens, Chinese-origin labs are collectively pushing 46%, and US labs’ share has fallen from roughly 70% to 30% year over year. Three Western flagship launches in one week did not move that needle. They moved the price sheet.

The Takeaway

If there’s one thing to carry out of this week, it’s this: stop reading the pricing page and start reading your invoice.

The intelligence race is basically a tie at the top. Four models within six points. In a tie, the tiebreaker is cost, and cost is not the number on the API docs. It’s tokens-per-task times price-per-token, and the two frontier models with the flashiest benchmarks (Fable 5, GPT-5.6 Sol) are also the ones that think the longest and bill the most per finished job. Grok 4.5 winning on cost-per-task while sitting fourth on intelligence is the entire lesson in one data point.

So here’s what I’m actually doing. I’m running Grok 4.5 for coding tasks where I can eyeball the output, precisely because it’s cheap per task, and watching it like a hawk because of that hallucination number. I’m keeping Gemini 3.5 Flash on math and creative work because paying 5x more for a rating rounding error is stupid. I’m letting the GPT-5.6 hype cool for a couple weeks until there’s lived-experience data instead of launch-day benchmarks. And I’m double-checking which model my tools are actually calling, because apparently that’s not a settled question anymore.

Rent everything, own nothing, read your bills. That’s where the model world sits in July 2026. See you next week, when Gemini 3.5 Pro and DeepSeek V4 have presumably set something on fire.