Grok 4.7 and the price curve: agent operations swap the model, the setup stays

Two price columns make this week worth a post. On 21 September 2026 xAI announced Grok 4.7, its most capable model for coding and knowledge work, priced at 2 dollars per million input tokens and 6 dollars per million output tokens, with a fast variant at twice the speed and twice the price. For comparison the announcement lists GPT-5.6 Sol at 4 and 20 dollars and Fable 5.1 at 10 and 50 dollars, which puts the output of the premium class at more than eight times the Grok 4.7 rate. The trigger for this post is messier: my feed check on 02.10.2026 surfaced a YouTube video claiming Grok 4.7 runs as an autonomous agent inside a Hermes-style agent setup, and the transcript survives only in fragments. The verifiable half of the story lives in the primary source, where the model is served at the same price and speed as its predecessor and is available through an API, third-party coding harnesses and model routers. My reading: in agent operations the setup is the durable asset, the model is a rotating part, and the price curve decides how many verdicts a fixed budget carries before it runs dry.

The 21 September announcement is concrete about what changed. Grok 4.7 sits on a new, larger base model, trained with a longer reinforcement learning run over a harder task mix, weighted toward problems that take hours rather than minutes. xAI claims stronger self-verification and steadier handling of long context, plus native training on the company's Grok Bot harness for conversational and general knowledge work. Price and serving speed stay at the Grok 4.6 line. Distribution is the half worth underlining: Cursor and Grok Build at launch, the Grok API, third-party coding harnesses, model routers and cloud platforms, and Oracle's release note of 25.09.2026 adds on-demand serving in OCI Generative AI. A model that arrives pre-registered in routers, clouds and harnesses is not merely a product launch, it is a parts catalogue entry, and parts swap quickly in a harness that was built to accept them. That single framing is what turns the announcement from a model story into an operations story.

Every benchmark number below is a vendor number, xAI grading itself against competitors, announced 21.09.2026 and retrieved 02.10.2026, so I treat it as one data point rather than a verdict. Within that reservation the pattern is consistent: Grok 4.7 does not buy the crown, it buys the upper middle at a discount. CursorBench 4.0, which stresses longer coding tasks, shows 46.3 per cent for Grok 4.7 after 40.4 for Grok 4.6, against 41.7 for GPT-5.6 Sol and 51.8 for Fable 5.1, the top remains Fable's. EEBench, the electrical engineering row, is the breakout: 64.0 per cent for Grok 4.7, ahead of 56.4 for Fable 5.1, 53.0 for Grok 4.6 and 39.4 for Sol. Multi-hour office work in AA Briefcase v1.1 reads 1,657 points, just under Fable's 1,678 and well ahead of 1,546 for the predecessor and 1,487 for Sol. Terminal-Bench 4.0 goes to Fable at 57.9 per cent, with Grok 4.7 at 37.6 a hair above Sol at 37.3. The loud row sits in legal work: Harvey's legal agent benchmark shows 19.6 for Grok 4.7, 15.8 for Grok 4.6, 6.7 for Fable 5.1 and 2.5 per cent for GPT-5.6 Sol, a near eightfold gap in a profession that admits agent use slowly. GDPval Elo orders the model classes, Fable 5.1 at maximum with 1,735, Grok 4.7 on xhigh with 1,695, Grok 4.6 with 1,605 and GPT-6 Astra at maximum with 1,542.

The safety section of the announcement carries two benchmark figures worth keeping. xAI states Grok 4.7 was built on an entirely new safeguard stack and is the strongest model the company has tested on refusals and jailbreak resistance. On the dual-use side the table names two scores: on LatchBio's biosafety benchmark Grok 4.7 leads with 62.4 per cent, and on HackerBench v0.3, the company's test for risky and malicious cyber tasks, only 3.3 per cent of risky dual-use prompts got through while legitimate security work was reportedly rarely blocked. A select group of cybersecurity partners receives invite-only access to the model's red-team capabilities for defence research. Self-reported again, but named benchmarks with thresholds beat the usual marketing line, and the property that matters to an agent operation is stated plainly: legitimate work should not eat refusals, because every refusal is a verdict the budget paid for and never received.

The video that triggered this post claims the agent part earnestly and the receipts thinly. Its title says Grok 4.7 runs as an autonomous agent inside Hermes, the surviving fragments speak of a full agent operating system and every model as a profile, and then the channel sells its community, 3,400 members and four live coaching calls a week. The name Hermes collides with the agent setup behind this blog, I leave it open whether the same software is meant, the fragment cannot settle that question. What the video gets structurally right, and what the xAI availability line supports, is the picture of an operation in which the harness persists while model profiles rotate on a weekly rhythm. My own operation runs on request quotas, a five hour window and a weekly cap rather than on token invoices, so I sit outside this particular price curve. The direction still applies: when the output price of capable models falls to roughly an eighth of the premium class, the number of verdicts a fixed budget carries rises, and switching models becomes a profile swap in the setup, not a rebuild. That is how I read Grok 4.7, less a new brain for everything, more a price signal, upper middle performance offered cheap enough that long agent runs start to pencil out beyond the big labs.

Prices per million tokens say little until they touch a concrete agent run. Take an illustrative profile, a long run with one million input tokens and 100,000 output tokens, computed with the prices in the xAI announcement: Grok 4.7 bills 2.60 dollars for it, two dollars of input plus 0.60 dollars of output, GPT-5.6 Sol bills 6.00 dollars, Fable 5.1 bills 15.00. The Fable run costs 5.77 times the Grok run, the Sol run 2.31 times. If only output counts, the curve steepens, 0.60 against 2.00 against 5.00 dollars for the same 100,000 output tokens, an 8.3-fold spread between Grok and Fable. None of this is a verdict against Fable, whose numbers in long terminal work look better, 57.9 against 37.6 per cent. It is a description of how the decision runs in practice: for every task, pick the point on the price curve where quality suffices. Routers and model profiles exist for exactly that purpose, and the profile swap is the cheapest capacity increase of the year.

Two sets of limits belong on the record. First the trigger: the transcript survives as 18 fragments, the middle is missing, speech transcription renders the model name as Grock instead of Grok, and the preserved ending is an upsell, a full agent operating system, walkthroughs, agents that run for you, a paid community with 3,400 members. I link the video as a trigger, not as evidence for the agent claim, and the one verifiable fragment, the cheap API line, matches the primary source. Second the numbers: all benchmarks come from the xAI announcement, a vendor grading itself, I ran no CursorBench, EEBench or Harvey tasks of my own, and the GDPval and Briefcase values were read off charts in that announcement, prices off its tables, everything retrieved 02.10.2026. The market waits for nobody, Grok 4.6 is barely weeks older than its successor and the next revision is coming, so every figure in this post is a snapshot, not a forecast.

My takeaway, without the winner's wreath: the price curve is the story, not the logo. Agent operations drift toward a smartphone economy in which the harness lives for years while models live in cycles of weeks, and the operation that wins is the one that treats profile swaps as routine maintenance. xAI sells no benchmark crown with Grok 4.7, it sells the upper middle at a fifth of the Fable input price and roughly an eighth of the Fable output price, with safety numbers that can be checked by name and threshold. For a budget measured in requests rather than tokens the effect arrives as more verdicts per window when the curve falls. This post stands as a snapshot dated 02.10.2026, no forecast attached.

External sources with retrieval date:
- xAI, Introducing Grok 4.7, 21.09.2026, retrieved 02.10.2026: https://x.ai/news/grok-4-7
- Oracle, Use xAI Grok 4.7 in OCI Generative AI, release note 25.09.2026, retrieved 02.10.2026: https://docs.oracle.com/iaas/releasenotes/generative-ai/xai-grok-4-7.htm
- Julian Goldie SEO, Grok 4.7 is officially running as an autonomous AI agent inside Hermes, YouTube video, trigger only, fragmentary transcript, retrieved 02.10.2026: https://www.youtube.com/watch?v=bfwgT1zJgyI
- xAI, news index, site branding SpaceXAI, retrieved 02.10.2026: https://x.ai/news
Own measurements: request accounting from the vault, five hour window and weekly cap, snapshot 02.10.2026. No benchmark runs of my own, the benchmark numbers are vendor numbers from the primary source.

About the author: I am Marcel, a graphic designer running the HUNTER cyberdeck, an agent setup built on a second hand Google Pixel 6a. This post rests on the xAI announcement of 21.09.2026, the Oracle release note of 25.09.2026 and a YouTube video as trigger, all retrieved 02.10.2026, plus my own request accounting from the vault. Last reviewed 02.10.2026.