Moonshot's Kimi K3.1 Pro has been callable since October 23: 1,048,576 tokens of context, $2 per 1M input tokens, $10 per 1M outputs. Its Terminal-Bench 4.0 score is 65.7%. GPT-6.2 has nothing official, just gpt-6.2-astra and gpt-6.2-sol strings in Google Cloud's quota console and an Information report hinting at a Friday launch. Its leaked prices are per finished task, not per token. Only Kimi can take requests today; GPT-stack teams should wait for a price sheet.
First question: what unit are you actually buying?
Start with the invoice, not the benchmark. Kimi meters tokens, so traffic models in a spreadsheet; the GPT-6.2 leaks price a finished agent run, scaled by effort. Only the first can serve live traffic today.
What the GPT-6.2 record actually contains
As of November 6: no model card, token rate, context figure, or benchmark table. The trail: quota pages in Google Cloud listing base_model:gpt-6.2-sol and base_model:gpt-6.2-astra, plus an Information report saying Astra, Sol, and a security-focused model could appear as early as Friday, November 6.
Three products emerge from those leaks: Astra, the volume tier (high-frequency work at lower cost, five effort settings, Low through Max); Sol, below OpenAI's strongest internal model but not a plain distillation; and an expensive, single-purpose security model. Leaked BenchCAD-style figures: Astra Low near $0.44 per task at 14%, Astra Max $11.05 at 59%, the security model roughly $11 to $37 at 76% to 88%. None is official; treat it as a preview of agent pricing, not a rate card.
What Kimi K3.1 Pro puts on the table
This half is verifiable. Moonshot shipped K3.1 Pro on Oct 23, 2026, as kimi-k3.1-pro and kimi-k3.1-pro-highspeed: $2 in, $10 out per million tokens, $0.20 cached; highspeed doubles to $4 and $20. Context is 1,048,576 tokens, though Kimi Code reserves the full window for higher-tier members. Requests take text and images; replies are text. Moonshot promised complete weights by November 4; Artificial Analysis tags it open-weight.
Scores: 65.7% on Terminal-Bench 4.0 versus GPT-6 Astra's 58.2% in Kimi's launch post; BenchCAD with tools flips it (92.7 Kimi, 95.9 Astra). The hallucination check matters more: on the Kimi Hallu Bench Hard test, where each claim is graded and lower wins, Kimi posts 8.3% against Astra's 5.5%. Artificial Analysis puts Kimi at 46 on its Intelligence Index, tasks averaging 35.9 minutes, so not low-latency. Every GPT-6 figure here describes GPT-6 Astra; GPT-6.2 has published no benchmark.
One workload, two price sheets
Same job: feed 1M input tokens, take 200,000 output, Kimi bills $4. At GPT-6 Astra's list rates ($10 in, $50 out) the same work costs $20, five times more; no GPT-6.2 token rate exists, so Astra is the only OpenAI reference.
Context is nearly a tie: 1,048,576 against 1.1 million, GPT-6.2 unknown. Either confirmed model handles repository review or long documents; the difference is what filling the window costs.
| Kimi K3.1 Pro | GPT-6.2 (leaks only) | |
|---|---|---|
| Pricing | $2 in / $10 out per 1M; cached $0.20 | None public; leaks quote per-task |
| Context window | 1,048,576 tokens | Not public |
| Status | Live since October 23, 2026 | Undocumented as of Nov 6, 2026 |
| API identifiers | kimi-k3.1-pro, kimi-k3.1-pro-highspeed | gpt-6.2-astra, gpt-6.2-sol (console strings) |
| Effort control | reasoning_effort: low / high / max; thinking always on | Five tiers reported for Astra |
| Modalities | Text + image in, text out | Not public |
| Tool surface | Function calling; documented web_search; REST search/fetch | Responses API built-ins; unconfirmed for 6.2 |
| Best for | Teams wanting published rates and an API they can carry | Teams already on OpenAI's agent stack |
Want Kimi without watching the token meter? It is one of eight models in our Multi-Model Coding Plan: one private key, shared points, the same low/high/max control (thinking stays on), and setup guides for Claude Code and Codex.
Tool access: build your own loop or rent one
Kimi ships web_search as a documented function, not a hidden $web_search: call it through function calling, or hit POST /v1/tools/fetch, /v1/tools/search_pro, /v1/tools/search. Server-side search also runs inside the Responses API; calls follow JSON Schema with tool_calls and tool_choice, support streaming and dynamic tool loading, and the Messages API speaks Anthropic's format. Other tools: memory, spreadsheets, conversion, dates, fetching. Kimi Code adds browser automation, MCP, skills, hooks, subagents.
OpenAI keeps most of the loop inside the response. The Responses API packs in function calling; a web_search tool with domain filters plus a token budget; remote MCP; shell access; computer use; file_search; a code interpreter. Chat Completions exposes web_search_options only on search-specialized models; Codex adds MCP, computer use, browser, web search, skills.
Kimi hands you parts and you own the loop; the Responses API rents the whole loop, with files, search, shell, browser handled inside the response. Its formats still travel between chat-completion clients. Thinking never switches off, which costs latency and tokens on easy prompts.
The cyber report: per-task billing at full scale
A November 5 OpenAI report describes an autonomous vulnerability-research pipeline built on an upcoming cyber model, run against the Linux kernel. The funnel: 40M lines of kernel code in, 5,034 audit targets out, then 2,308 candidate findings, 1,724 reproduced vulnerabilities, 291 working leak PoCs, 1,169 exploit programs, 948 local privilege escalation bugs with patches. An LPE gives an ordinary user root. The run burned $1.213M in API credits, about $1,279 per confirmed LPE. In one Google kernelCTF exercise it flagged a flaw in the test VM itself, then wrote an exploit with no QEMU source.
The report never says GPT-6.2 Cyber; it credits an unnamed upcoming model. It points at orchestrated models for long jobs where one run costs thousands yet beats weeks of expert time. In a 166-exploit sample, agents reached for at least 10 techniques outside experts had never seen. Human technique lists lifted GPT-5.5 Cyber from 76.2% to 88.9%; GPT-5.6 Cyber did better with no hints. This is a specialist product: Kimi sells general capability at token rates, the cyber line sells finished security work per task.
Picking a side
Building something now? Kimi K3.1 Pro fits as the default: published rates, portable API, room for large repos and long documents. Caveats: it trails the fastest models on speed, and it lost the hallucination test to GPT-6 Astra. Keep automated checks and human review.
On Codex, the Responses API, or computer use? Wait for official GPT-6.2 numbers. Those leaked tiers price a completed agent run, not raw tokens; if that matches your workflow, the higher per-job cost may be worth it. Do not move production on a rumor.
Security research: watch the cyber line, but it does not replace a general coding assistant. Treat it as a pipeline component with its own budget.