On OrcaRouter the grok api page lists five models, and one of those rows wears two labels at once. `grok/grok-4.3` is the model the page’s header names under cheapest input at $1.25 per million tokens, and it is also the row sitting at 10 on our own 0-to-10 quality composite — a composite score we build ourselves from the quality dimensions we grade each model on, not a third-party benchmark. Across the seven provider catalogues on the site, xAI is the only vendor where the flagship and the cheapest row are the same SKU, and the page’s own cost answer names it twice inside one sentence.
That repetition is not a slip. Read the sentence as a whole and the shape of the entire catalogue falls out of it: five models, four of them priced per token, and an input column that barely moves. There is no cheap tier to graduate down to and no expensive tier held back for special occasions. There is one price point, and the question a buyer actually has to answer is not which row is cheapest but whether a single price point is enough of a line-up to route real traffic against.
The cost answer that names one model twice
The Grok page’s cost FAQ reads: “grok/grok-4.3 is $1.25 per million input tokens and $2.5 per million output tokens; the cheapest xAI model here is grok/grok-4.3 at $1.25 per million input.” Parsed as prose that is an odd sentence — the answer to “what does this cost” and the answer to “what is cheapest here” are the same model, so the second clause sounds like it is repeating the first by mistake. It is not a mistake; it is the catalogue describing itself. On every other provider page the two clauses name different rows, because those catalogues have a spread wide enough for the flagship and the bargain to be separate objects. Here they collapse into one.
For anyone shortlisting on price, that collapse changes the job. The usual first pass — sort by input price, then check whether the cheap row is good enough — returns a single row, and the sort has nowhere left to go. The interesting work moves from the price column to the other columns, because price is no longer a variable you can trade against quality. At xAI, this generation, they are the same decision.
Five rows, one price point
| Model | Input / 1M | Output / 1M | Context | Max output | Quality (ours) |
| `grok/grok-4.3` | $1.25 | $2.50 | 1M | — | 10 |
| Grok 4.5 | $2.00 | $6.00 | 500K | — | 9 |
| Grok 4.6 | $2.00 | $6.00 | 500K | — | 9 |
| Grok 4.7 | $2.00 | $6.00 | 500K | 450K | 9 |
| `grok/grok-imagine-image` | $0.200 per request | — | — | — | 5 |
The quality column is our composite, not a vendor claim and not a benchmark score. The prices are the vendor’s published rates, and OrcaRouter adds no markup, so the table is what calling xAI directly would bill you.
Four of the five rows are token-priced, and every one of them lands on either $1.25 or $2.00 on input. The whole input column spans 1.6x, from $1.25 to $2.00, and the output column spans 2.4x, from $2.50 to $6.00. Put that next to the other catalogues on the same site and the contrast is the point: the OpenAI page on OrcaRouter runs from a $0.02 embedding row to a $30.00 flagship on input, and the Anthropic page from $1.00 to $10.00. Those are pages where a price decision and a capability decision are separate exercises. Here they are one exercise, and the only row that is meaningfully cheaper than the flagship is the flagship.
Notice also what the cheap end of the column is not doing. There is no small, fast, sacrifice-everything row at a fraction of the flagship price — no sub-dollar input row to serve the high-volume, low-stakes traffic that almost every production stack has somewhere. The cheapest token-priced model on the page is the most capable one. If your workload mix includes a noisy summarisation pass or a classification sweep that a small model would handle perfectly, that pass still bills at flagship rates.
The image model, billed per request
The fifth row is different in kind, not degree. `grok/grok-imagine-image` is billed at $0.200 per request, and its token columns are simply not part of the deal: there is no input price and no output price, because the unit of sale is the request. Its composite score is 5 on our scale, which is what you would expect from an image row measured against text models on our text-weighted dimensions.
Per-request pricing is worth separating from the two things it is often confused with. It is not free tokens — a request costs $0.200 whether the response is long or short — and it is not a token rate you can multiply by a token budget. If you forecast spend the way you forecast for the four text rows, the image row will break the model, because its cost scales with the number of calls rather than the size of them. That makes cost control for image generation a question of call volume and batching, not prompt length.
What a one-price-point catalogue does to a shortlist
The practical consequence is a shortlist with one entry on it, and the discipline that follows is different from the usual one. Normal multi-provider shortlisting asks which model is good enough at the lowest rate. Here there is no rate ladder to walk down, so the questions invert. Does the flagship do what you need? If yes, you are done: you buy it and you do not spend a week of engineering time benchmarking a cheaper alternative, because there is no cheaper alternative. If no, no lower row rescues you — 4.5, 4.6 and 4.7 cost half again as much on input and score one point lower on our composite, so a capability gap at the top of the table cannot be closed by shopping downwards.
That is a real simplification, and it is not universally good news. A catalogue with one price point is cheap to reason about and expensive to optimise. There is no negotiating lever, no second tier to move batch traffic onto, and no cheap model to hedge a risky new workload behind. What you get in exchange is honesty about what the lineup is: a small, tightly priced set of frontier models plus one image SKU, with the price column deliberately not used as a product-segmentation tool.
Where the capability row matters more than price
With price nearly constant, the capability census on the page does more of the work than the price column. All five rows carry vision — five of five — while tools and reasoning come in at four of five, and JSON mode at five of five. That is an unusual profile: the multimodal flag is the one nothing is missing, and the agentic flags are the ones with a gap. For a pipeline that needs structured JSON back from a model, every row qualifies. For a pipeline that needs the model to call tools, one row does not, and with a price column this flat the capability column is the only filter worth applying.
The page also reports a median p50 of 5,000 milliseconds across the catalogue. That is the number to hold against your interactivity budget, and it is the one dimension where a one-price-point catalogue gives you no room to shop: you cannot buy latency back with a cheaper trade, only accept or reject the tier as a whole.
The takeaway
One vendor, one price point, and a flagship that doubles as the budget row is a small piece of information that carries a lot of consequence. It tells you the lineup is short by design, that price is not being used to segment buyers, and that a cost problem in this catalogue is a workload-shape problem rather than a model-selection problem. It also tells you what to check first: not which row is cheapest, since that is settled, but whether the row you are already going to buy clears your capability bar on tools, context and the p50 median you are buying into.
A catalogue that spans 1,500x from its cheapest to its dearest row is a menu. A catalogue that spans 1.6x is a product. The xAI page is the second kind, and reading it as the first kind is how teams end up benchmarking a row that was always going to be the one they picked.
Sourcing note: model count, cheapest-input statistic, median p50, the pricing rows, the context and max-output figures and the capability census are from the OrcaRouter Grok provider page read 2026-10-14; the quality column is our own 0-to-10 composite built from our own quality dimensions, not a third-party benchmark, and every price is the vendor’s published rate [VENDOR] with $0 markup. The cross-provider price ranges quoted for the OpenAI and Anthropic catalogues are from their OrcaRouter provider pages on the same date [OURS]. The per-request image rate is the vendor’s published rate [VENDOR]. All figures read 2026-10-14.
