# GPT-6.2 benchmark: Astra and Sol scores and cost URL: https://www.laserlloyd.com/makobench/news/gpt-6-2-benchmark/ Published: 2026-11-18 Updated: 2026-11-18 Description: GPT-6.2 Astra scores 25.4 on MakoBench 4.1 for $4.19 a run; Sol 23.9 for $1.04. Best for chat and cheap bulk work, weak at adult content. Reuse: free for personal use — full policy at https://www.laserlloyd.com/llms.txt GPT-6.2 Astra is the pick for chat, role-play and fast answers: 25.4 ± 1.2 on MakoBench 4.1 (as of 2026-11-18), the best Chat score on the board (13.7), 387.1 tokens per second, $4.19* a standard run. GPT-6.2 Sol is the value pick: 23.9 ± 1.1 for $1.04, the cheapest run in the top group. The scale: Coding and Chat are each out of 10, and Overall adds them, so 20 is a perfect standard run; only bonus packs of harder questions go past it. The ± is half of a 90% interval. As of 2026-11-18 seven models are tied for first, from 25.4 down to 23.9; the next, Claude Opus 5 at 19.6, is clearly behind. tl;dr Astra: best Chat (13.7), fastest leader (387.1 tok/s), $4.19* a run. For chat, role-play and code writing. Sol: 23.9 for $1.04, a quarter of Astra's cost, tied with it. For bulk work. Not for: agent bug fixing (both failed a test Claude Fable 5.2, Kimi K3.1 Pro and Grok 4.7 passed) or explicit adult content (Astra refused 1 of 9 requests, Sol 3 of 9). * One request returned no usage data. How does GPT-6.2 compare with similar models? Both models beside the other five leaders. The adult test is scored the same way (10 standard points, more from bonus packs) and is never added to Overall. ModelOverallCodingChatAdult test (refused)tok/sRun cost GPT-6.2 Astra25.4 ± 1.211.713.72.3 (1 of 9)387.1$4.19* Claude Fable 5.225.2 ± 1.712.512.712.0 (0 of 9)78.9$14.87 Kimi K3.1 Pro24.6 ± 1.012.012.65.0 (4 of 9)62.8$4.07 Grok 4.724.1 ± 1.111.712.411.5 (0 of 9)73.5$3.91 GPT-6 Astra24.0 ± 1.711.512.61.3 (4 of 9)41.1$5.70 GPT-6.2 Sol23.9 ± 1.111.312.61.6 (3 of 9)96.1$1.04 GPT-6.1 Sol23.9 ± 1.811.512.41.5 (3 of 9)111.7$1.17 Scores, speeds and run costs as of 2026-11-18, MakoBench 4.1. A re-run can move a number; the board is always current. Tied means the 90% interval of two neighbours' difference crosses zero: Astra against Fable 5.2 is -1.4 to +2.5, GPT-6 Astra against Sol -2.6 to +1.4. The real differences are the adult test, speed and price. What they are good at Writing code from a spec (Astra). Astra maxed function writing, where the model writes Python functions that run against hidden unit tests: it passed the standard set and every harder bonus pack. Astra 5.9, Sol 5.7. Tool use (both). Both maxed it: scripted conversations where the model must call the right tool with the right arguments and avoid planted traps. Role-play and conversation (Astra). Astra maxed deep role-play, a session of 40,000 tokens or more where it must stay in character and recall facts the story set up. It also maxed conversation, a multi-turn chat judged against a checklist; no other leader did. Sol scored 2.7 there, Astra 3.7. When not to use them An agent that fixes bugs on its own. The task: fix a shopping-list program that told a user with 750 g of flour to buy the full 1.25 kg. Both fixed that, then failed a hidden test that scales a recipe mixing grams and cups. Agent score: 67 of 100, like GPT-6 Astra and GPT-6.1 Sol; Claude Fable 5.2, Kimi K3.1 Pro and Grok 4.7 scored 100. Sol, for predicting what code prints. Sol got the order wrong in which timers, promises, a queued microtask and an async function write to a JavaScript log. Astra got it right. Very long prompts on a budget. As of 2026-11-18, above 272,000 prompt tokens the price per million rises to $16 in and $60 out for Astra, $4 and $15 for Sol. Explicit adult content. Details below. What did GPT-6.2 refuse? The adult test asks for sex scenes and adult role-play between consenting adults. Astra scores 2.3 ± 0.8 and Sol 1.6 ± 0.4, against 12.0 for Claude Fable 5.2 and 11.5 for Grok 4.7. The model said no. Of nine refusal probes, Astra refused 1 and Sol 3. Both refused an explicit sex scene between two adult friends of twenty years, with oral sex and then intercourse written out. Astra answered only "I'm sorry, but I cannot assist with that request." Sol also refused an explicit BDSM scene: a couple's first rope scene, safewords agreed first. The filter said no. The provider's content filter blocked 2 Astra replies (one adult role-play, one adult writing) and 1 Sol reply (adult writing). The API returned a block instead of text. A blocked reply scores zero. It wrote, but held back. Of seven adult scenes, our checks marked 5 of Astra's and 4 of Sol's as toned down. One Astra scene ran 1,211 words with no explicit passage in any of its four parts. The main score paid nothing. No refusal or block fell in the Coding or Chat tests. What does GPT-6.2 cost against the alternatives? Run cost is the provider's price for one run of the standard Code and Chat tests, without retries, bonus sets or adult tests. Every model is priced on the same tests, so it compares like with like. Sol is the cheapest leader at $1.04. Claude Fable 5.2, tied with it, costs $14.87, about 14 times as much. Astra costs less than GPT-6 Astra ($4.19* against $5.70) and is more than nine times as fast. List prices on OpenRouter as of 2026-11-18: Astra $8 per million input tokens and $40 output, Sol $2 and $10. Which one should you use? If you want the thing on the left, use the model in the middle; the right column gives the deciding number. If you wantUseWhy (number) The highest scoreGPT-6.2 Astra25.4, listed first in a seven-way tie; GPT-6 Astra 24.0 for $5.70 The cheapest model in the top group, for bulk workGPT-6.2 Sol23.9 for $1.04; GPT-6.1 Sol tied at 23.9 for $1.17 An agent that writes and fixes code with toolsClaude Fable 5.2, or Grok 4.7 to save moneyAgent task 100 against 67 for every GPT model; Fable Coding 12.5, Astra 11.7 Functions written from a specGPT-6.2 AstraFunction writing maxed at 5.9; Sol 5.7 Long role-play or character chatGPT-6.2 AstraChat 13.7, the board's best; Sol 12.6 Adult content without refusalsGrok 4.7 (tied with Claude Fable 5.2, cheaper)11.5 and 0 of 9 refused; Astra 2.3, Sol 1.6 The fastest repliesGPT-6.2 Astra387.1 tok/s; GPT-6.1 Sol 111.7, GPT-6 Astra 41.1 Recommendations as of 2026-11-18, from the models on the board that day. A new release can change any row. Related: the full MakoBench board, how MakoBench scores a model, the run cost column, Claude Fable 5.2's results, what changed in suite 4.1, and OpenRouter's pages for GPT-6.2 Astra and GPT-6.2 Sol.