A Grok 4.7 request that reaches 200,000 prompt tokens triggers long-context rates for every token in the request. xAI's pricing table puts the short-context rates at $2 per million input tokens and $6 per million output tokens. At the threshold, those rates become $4 and $12. A request with 199,999 input tokens and 10,000 output tokens costs about $0.46. Add one input token, keep the output the same, and the bill becomes about $0.92.
That boundary is the developer story hiding behind a busy launch. Grok 4.7 reached 329 Hacker News points and 211 comments in less than four hours, according to the launch-day snapshot of the discussion. The attention shows that developers are watching. xAI's performance and safety figures still need to stand on their own evidence.
What xAI shipped
xAI describes Grok 4.7 as its most capable model for coding and knowledge work. The company says it trained a larger base model than Grok 4.6 with a longer reinforcement-learning run, using a harder task mix weighted toward work that can take hours. It also says the model checks its work more carefully, manages longer context better, and was trained to understand the Grok Bot harness. Those are company claims, not independent findings.
The model documentation fills in details that matter during integration. The API model name is grok-4.7. It accepts text and images, returns text, has a 500,000-token context window, and has a May 2026 knowledge cutoff. Developers can choose low, medium, high, or xhigh reasoning effort. High is the default. The model supports function calling, web and X search, code execution, the Responses API, and Chat Completions.
Availability is broad on launch day. xAI lists its own API and Grok Build, Cursor, model gateways including OpenRouter, Vercel, and Cloudflare, plus other coding harnesses and cloud platforms. That makes Grok 4.7 easy to trial without changing an entire agent stack. It also means invoices and behavior can differ by provider, so the model name alone does not identify the service conditions.
The 500K window has a price step
The advertised $2 input and $6 output prices are starting rates. At 200,000 prompt tokens, input rises to $4 per million, cached input to $1 instead of $0.50, and output to $12. The pricing rule applies the long-context rate to all tokens in that request. It does not charge the higher rate only on the portion above 200,000.
This matters most in agent loops, where tool results, file contents, and earlier turns accumulate in the prompt. A run that sends 300,000 input tokens and receives 50,000 output tokens costs $1.80 at the listed long-context rates. The same token counts evaluated at the short-context rates would cost $0.90. The 500,000-token capacity can save an application from truncating useful state, but using that capacity is a budget decision.
xAI's docs recommend setting prompt_cache_key for the Responses API or the x-grok-conv-id header for Chat Completions. The company says those values route a conversation to the same server and make cache hits more reliable. Without them, a request can land on a cache-cold server and pay the full input rate. The same page points long-running agents toward context compaction. Both controls become more valuable near the 200,000-token boundary.
Grok 4.7 Fast adds another tariff. It runs only in Cursor and Grok Build, outside the public xAI API. The pricing page lists $4 input and $12 output below 200,000 prompt tokens, then $6 and $18 above the boundary. xAI says the fast service provides twice the output speed. Teams comparing it with the standard API need to record which service produced each latency and cost result.
Reasoning effort is the second meter
xAI's CursorBench 4.0 chart makes the tradeoff unusually visible. At low effort, Grok 4.7 scored 33.1% at an average cost of $1.58 per task. Medium reached 41.6% at $3.49, high reached 43.9% at $4.69, and xhigh reached 46.3% at $6.01. These are benchmark averages from xAI's launch page, so they are not forecasts for an individual repository. They do show why an effort setting belongs in both evaluation records and production telemetry.
The curve is uneven. Moving from low to medium bought 8.5 percentage points for another $1.91 per benchmark task. Moving from medium to xhigh bought 4.7 points for another $2.52. That does not prove medium is the right setting for every job. It suggests that assigning xhigh to every request can spend more on the shallow end of the measured return curve. High, the API default, sits between those points.
An independent result points in the same direction, with a different test mix. Artificial Analysis scores Grok 4.7 xhigh at 46.45 on its Intelligence Index and high at 46.33, a gap of about 0.12 points. Its index combines ten evaluations, and the page had not yet posted measured speed or cost for the model when checked. That narrow gap cannot settle a workload-specific choice, but it is a reason to test high before paying for xhigh everywhere.
The benchmark profile is uneven
xAI's comparison table resists a single league-table reading. Grok 4.7 xhigh scored 46.3% on CursorBench 4.0, ahead of Grok 4.6 high at 40.4% and GPT-5.6 Sol max at 41.7%, while Fable 5.1 max reached 51.8%. On DeepSWE v1.1, Grok recorded 71.0% at high effort. That was above Grok 4.6's 65.2% and just below GPT-5.6 Sol's 72.7%.
Longer terminal work produced a wider spread. Grok 4.7 scored 38.0% on Terminal-Bench 4.0, nearly twice Grok 4.6's 20.3% and slightly above GPT-5.6 Sol's 37.3%. Fable 5.1 reached 57.9%. On EEBench, xAI reports 64.0% for Grok 4.7 against 53.0% for its predecessor. The release looks stronger on several kinds of work, yet the leading model changes with the test.
The comparison settings also differ. xAI used xhigh or high for Grok 4.7 depending on the row, high for Grok 4.6, and max settings for the two competing models. Those choices belong beside the scores because effort affects tokens, latency, and cost. A benchmark win at a larger inference budget may still be the right purchase, but the budget is part of the result.
Safety figures need their own test plan
xAI says Grok 4.7 uses an entirely new safeguard stack. The company reports 62.4% on LatchBio's biosafety benchmark and says only 3.3% of risky dual-use prompts passed its HackerBench v0.3 controls. It has also opened invite-only access to selected cybersecurity partners for defensive red-team research. The announcement supplies the outcome figures, though it does not place sample counts or confidence intervals beside them.
Teams handling security, medical, or biological material should therefore test two failure modes on their own task set: dangerous assistance getting through and legitimate work being refused. xAI claims progress on both, but a single aggregate score cannot tell a security team whether its incident-response prompts, internal terminology, or tool permissions behave acceptably. The check needs the same model version, effort setting, and harness intended for deployment.
Record the configuration, then compare
A useful trial can start small. Take a fixed set of real coding or knowledge-work tasks and run each at medium, high, and xhigh. Record completion quality, total input and output tokens, wall time, retries, and the final charge. Repeat long agent runs on both sides of the 200,000-token threshold, then repeat them with stable cache routing and context compaction. That reveals whether the larger window preserves useful state or merely carries expensive history.
Independent speed measurements, cache-hit behavior across providers, and real project results will matter more than another launch chart. Artificial Analysis had a quality score ready on release day but still showed speed and cost as unavailable. Cursor and Grok Build also expose a fast service that the public API does not, which makes provider labels necessary when users share results.
For every Grok 4.7 result, write down the effort level, prompt length, cache state, and endpoint beside the score. Without those fields, the result cannot tell another developer what the measured quality cost.