Grok 4.7 kept the price list. Your bill may not.

SpaceXAI’s September 21 post sells Grok 4.7 as “same price and speed as Grok 4.6.” The meter still starts at $2 per million input tokens and $6 per million output tokens. Cursor and Grok Build could select it the same day. Read only that sentence and the upgrade button looks cheap to press.

The catch is one layer down. Artificial Analysis scored the model the same day: 46 on the Intelligence Index, only two points above 4.6, while xhigh effort burned about 81,000 output tokens per Index task against roughly 36,000 for 4.6 at high and 38,000 for 4.6 at xhigh. The rate card did not move. Tokens per finished task did.

We covered the mid-year model pile-up and how feature pages diverge from intelligence boards. This piece adds one narrow claim: a same-price version bump is not a same-cost version bump until you lock effort and read the token line on the invoice.

What the vendor table is actually comparing

SpaceXAI’s main chart puts Grok 4.7 at xHigh, Grok 4.6 at High, and rivals at Max. Coursiv’s September 22 write-up puts it flatly: the gaps mix model quality with spend. Higher effort is supposed to spend more reasoning tokens. Pick a default model off that chart and you are watching a high-burn setting beat a lower-burn setting, then calling the gap a generation jump.

The published numbers are still worth keeping. CursorBench 4.0: 46.3% for 4.7 xHigh versus 40.4% for 4.6 High. DeepSWE v1.1: 71.0% (marked high effort on the post) versus 65.2%. EEBench 64% versus 53%. AA Briefcase 1657 versus 1546. Terminal-Bench 4.0 38% versus 20.3%. Harvey Legal 19.6% versus 15.8%. HealthBench Professional 56.7% versus 48.5%. Treat the legal and clinical columns as supplier slides. Rival scores on those two rows look oddly weak, which is a reason to distrust the frame.

A separate GDPval chart lists Elo: Fable 5.1 max at 1735, Grok 4.7 xhigh at 1695, Grok 4.6 high at 1605, GPT-6 Astra max at 1542. That is one of the few places Astra appears in the launch package. We already tracked Astra’s security-customer rollout. On that professional-document chart, Grok places itself ahead of Astra. On Artificial Analysis’s Intelligence Index, Astra still sits at 53 and Grok 4.7 at 46. Different charts answer different questions. Paste only one into a procurement email and the thread argues with itself.

What the independent run adds

Artificial Analysis re-ran with a shared harness. Intelligence Index 46 against 44 for 4.6. AA-Briefcase Elo 1657, up 111 from 4.6 high, sitting just behind Claude Opus 5 and Fable 5.1. GDPval-AA 1695, up 90 from 4.6 high. Analytical-quality Elo jumped from 1690 to 1994; presentation quality barely moved. Long-horizon agentic knowledge work is where 4.7 actually steps forward.

On the Coding Agent Index with first-party Grok Build, 4.7 xhigh scores 56, up nine points from 4.6 xhigh at 47. DeepSWE, Terminal-Bench, and SWE-Atlas-QnA all rise. That score is model-plus-harness, a different exam from the Intelligence Index, which standardizes the agent loop across vendors. If your workflow already lives inside Grok Build, the harness number matters more than the bare model line. If you mix agents inside Cursor, start with the shared harness, then read the branded page.

Elsewhere the Index is mostly flat. Some Terminal-Bench storytelling in the vendor post is louder than the AA breakdown, which describes broad parity with 4.6 outside the agentic-knowledge cluster, with small gains and small regressions. “Best coding Grok yet” does not mean “wins every column.”

Tokens: the line the rate card does not print

AA is blunt. Grok 4.7 xhigh uses about 81,000 output tokens per Intelligence Index task. Grok 4.6 high uses about 36,000. GPT-6 Astra max uses about 27,000. Across the full evaluation suite, 4.7 xhigh emitted roughly 240 million output tokens against about 94 million for 4.6 high; the suite median sits near 92 million. Coursiv repeats those figures and adds the matched-effort punchline: High and xHigh both score 46 on the Intelligence Index. Maxing effort did not move the composite score. It moved the token counter.

The arithmetic is dull, on purpose. Output still costs $6 per million tokens. If a task’s output climbs from 36,000 to 81,000 tokens, that slice of the bill scales by about 2.25×. Input still exists. Cache hits still exist. On long agent runs, output is often the fat line. SpaceXAI can say the price is unchanged. AA can say the same sticker buys fewer completed tasks. Those two sentences can both be true this week.

Cursor’s model page adds another meter. Standard context is 256k; long context stretches to 500k. Once input crosses 256k, standard requests bill at 2× those rates and Fast requests bill at 3× the standard rates. Effort levels are low, medium, high, and xhigh, with high as the named default. Fast is $4 / $12. Click “faster” in the IDE and you meet the second row of the price table before you notice you left the first.

AA also clocks answer speed around 188 tokens per second on long prompts and about 7.1 minutes per Intelligence Index task at xhigh. Latency is not free either. A model that thinks longer can still be the right tool — just not the one implied by “same speed as 4.6” on a launch headline.

Safety claims and missing pages

The launch post describes a new safeguard stack and calls 4.7 the company’s strongest result yet on refusals and jailbreak resistance. LatchBio’s biosafety benchmark is cited at 62.4%. On HackerBench, about 3.3% of risky dual-use prompts slip through, with a claim that legitimate security work is rarely blocked. Red-team capabilities are invite-only for selected defense partners. Those are vendor numbers, not a third-party retest you can paste into an audit packet without a caveat.

Coursiv lists what the post skips: no ship date for grok.com, mobile apps, or SuperGrok plans; no parameter count or training disclosure; image input is listed without a public vision benchmark; no link to a system card. For a buyer, missing pages are a signal. API and Cursor availability is not a consumer cutover.

How this hooks to older SynapseWire pieces

The July model-wave piece was about choice cost when several frontiers land together. This time the choice cost is sharper: a 0.1 version bump from the same vendor, unchanged list rates, and a default effort nudge that can reshape the invoice. The Claude-versus-Gemini article was about feature pages drifting away from intelligence boards. Here the drift is between the rate card and the token ledger. The Astra piece was about rollout sequencing. Grok 4.7’s sequence is developer surfaces first, consumer surfaces unannounced.

The method that survives contact with finance is boring: freeze a batch of your own tasks, run 4.6 and 4.7 at the same effort, score quality and output tokens together. The vendor’s main table answers a marketing question; your CI logs answer an accounting question.

Where Grok 4.7 actually moves the frontier

Strip away the price rhetoric and the independent story is narrower than the launch tone. Agentic knowledge work moved. AA-Briefcase and GDPval-AA show clear Elo gains, with analytical quality doing most of the lifting. Coding agent performance moved when the model rides Grok Build. Raw Intelligence Index movement is modest. Several non-agent Index tasks look like noise around 4.6.

That shape fits a lab optimizing for multi-hour tool loops and document production, not for a two-point composite brag. It also fits a go-to-market that opens Cursor and Grok Build on day one and leaves consumer chat unannounced. If you live in those surfaces, the release is aimed at you. If you live in a phone app, the release has not told you when you get it.

Hallucination metrics in the AA note are mildly better: AA-Omniscience hallucination rate 29% versus 34% for 4.6 high, accuracy roughly flat, composite Omniscience Index 32 versus 30. Mild is doing a lot of work there. Do not rewrite your citation policy for a five-point hallucination dip on one private suite.

Procurement language that survives contact with finance

If someone forwards the launch link and asks whether to flip the org default, answer with four lines. One: list rates for the first tier match 4.6. Two: vendor charts compare unmatched effort. Three: independent Index gain is about two points with roughly double the output tokens at xhigh. Four: we will decide after a fixed eval at matched effort, with token totals attached.

That answer is boring enough to survive a finance review, and it still leaves room for 4.7 to win. Nothing here says the model is worse. It says the marketing sentence and the cost sentence are different sentences, and conflating them is how teams “save money” into a larger bill.

What to watch this week

Watch per-request tokens in Cursor or the API console, not only pass/fail. Watch whether Grok Build jobs get longer and more expensive before they get better. Watch for a SpaceXAI chart that finally aligns effort levels. Watch for a quiet consumer rollout. Watch whether AA or anyone else publishes a fuller matched-effort curve beyond the Intelligence Index composite.

The sticker price really did not rise. Whether each job got cheaper is a question for your usage export. Hearing “same price” as “same cost” is the most expensive sentence in this release.