---
title: "GPT-6.2 Astra and Sol: pricing, benchmarks, and flat coding scores"
description: "OpenAI released GPT-6.2 Astra and GPT-6.2 Sol: Astra drops to $8 / $40 per million tokens and hits 78.3% on Agents' Last Exam at $2.89 per task. On Artificial Analysis's Coding Agent Index, though, Astra ties GPT-6 Astra, Sol drops 3 points, and Astra's hallucination rate climbs from 50% to 59%. Official benchmarks, third-party results, hands-on tests, and task-level bills."
url: https://codingpass.ai/blog/gpt-6-2-astra-sol-cheaper-not-smarter
language: en
updated: 2026-11-18
---
# GPT-6.2: cheaper per task, no better at code

Every chart in OpenAI's launch post plots score against cost per task. The savings are real: Astra runs Agents' Last Exam for 5% of Claude Fable 5.2's task cost. Yet coding scores didn't budge, hallucinations climbed, and testers still complain about unreadable plans. Here is how the official numbers, Artificial Analysis data, and developer feedback line up.

[CodingPass team](https://codingpass.ai/about) Published Nov 18, 2026 Updated Nov 18, 2026 10 min read

OpenAI has released GPT-6.2 Astra and GPT-6.2 Sol. Astra is the new flagship, pitched for "the hardest projects," while Sol serves as the everyday workhorse. Both shipped simultaneously across ChatGPT, Codex, and the API under the IDs `gpt-6.2-astra` and `gpt-6.2-sol`.

Testers had been spotting GPT-6.2 Sol in ChatGPT A/B rollouts since early November, nearly 20 days before the official launch. What arrived has one clear pitch: lower task bills. Nearly every chart on OpenAI's launch page puts cost per task on the horizontal axis, backed by a 20% list-price cut on Astra and task costs that undercut rivals by multiples.

Put the launch page next to Artificial Analysis (AA) benchmarks and developer testing, though, and the trade-off is sharp. Coding capability stalled, hallucination rates went up, and planning and communication remain weak spots. By day two, developers in the community were already switching back to Claude.

**Bottom line:** GPT-6.2 genuinely cuts task costs, and its gains are concentrated in scientific research, 3D, and legal documents. On AA's Coding Agent Index, Astra ties GPT-6 Astra at 53 while Sol drops 3 points from GPT-6.1 Sol, and Astra's hallucination rate rises from 50% to 59%. Developers running long agentic jobs report models spinning in place and struggling to summarize what they did. Keep Claude for writing code, and switch **code review** to GPT-6.2 Sol.

## Early reception: developers push back

OpenAI's launch charts show clean cost curves, but developer sentiment on Hacker News and X turned negative within 24 hours. Complaints clustered around a few recurring failures.

**Stalling on long-running tasks.** The sharpest pushback came from T3 Code creator Theo, who published a letter he had sent OpenAI during early testing on November 17 under the headline "Why I hate GPT-6.2 Astra"—drawing over 1.75 million views in a day. Testing on a project rewriting the TypeScript compiler in Rust, where earlier OpenAI models had burned over $300,000 in tokens without progress before Opus finished the job in under two weeks for roughly $20,000, Astra "ran for 7 days and landed 0 improvements." It fixated on shrinking binary size and slowed the compiler down by 56%; redirected to focus on performance, it ran for another four days with zero changes. On a separate run, two hours in, it stalled itself by stopping to ask what to name an internal intermediate file.

**Unreadable summaries.** That same opacity shows up in planning tests. Theo now routinely hands completed Astra runs to Opus 5.5 just to find out what happened. When he asked why a worktree on one machine was slower than three others, Astra wrote four paragraphs on pnpm caching; Opus 5.5 replied in one line: "it's because the ssd is slow." Changing the system prompt to ask for simpler explanations didn't fix it.

**Heavy oversight.** On Hacker News, developers noted that Astra is capable but "needs SO much oversight and careful prompting that Opus 5.5 doesn't need." Others called Sol prone to overengineering and questioned whether Astra's code quality justifies its price. Code review findings can also be inflated: one developer had Astra flag 12 issues in production code—8 of them P2—only to see the list shrink to 8 P3s when cross-checked alongside Fable 5.2.

**No faster than before.** Measured throughput on both models sits at 20–80 tokens per second, unchanged from the previous generation and around 5x slower than Anthropic's models. Several developers wrote that OpenAI has had a rough two months and that they had already moved back to Opus 5.5.

Praise did land where the benchmarks are strongest: Theo acknowledged that Astra is a strong code reviewer, world-class at 3D and games, and solid at computer use. Some HN commenters preferred Sol over Opus 5.5, noting that its UI designs finally caught up with Claude.

Where benchmarks and hands-on experience diverge is exactly what benchmarks rarely measure: whether an agent gets stuck mid-run and whether it can explain its work clearly afterward. Cost-per-task curves count one run, not the human time spent babysitting the model or paying a second model to translate its output.

## Models, API pricing, and fine print

Astra's input and output prices both drop 20%, and cached input falls from $1 to $0.40 per million tokens. Sol's pricing is unchanged. At $8 / $40 per million tokens, Astra matches Fable 5.2's reduced list price, just as GPT-6 Astra matched Fable 5 before it.

| Model | Input | Cached input | Cache writes | Output | Knowledge cutoff |
| --- | --- | --- | --- | --- | --- |
| GPT-6.2 Astra | $8 | $0.40 | $10 | $40 | 2026-05-18 |
| GPT-6 Astra | $10 | $1 | $12.50 | $50 | 2026-04-30 |
| GPT-6.2 Sol | $2 | $0.10 | $2.50 | $10 | 2026-06-01 |
| GPT-6.1 Sol | $2 | $0.10 | $2.50 | $10 | 2026-04-30 |

Both models support a 1,050,000-token context window and up to 128,000 output tokens. The fine print in OpenAI's [model documentation](https://developers.openai.com/api/docs/models/gpt-6.2-astra) matters for long agent sessions:

- **Cache writes cost extra.** A 30-minute cache write on Astra is $10 per million tokens—25% more than standard input (down from $12.50 on GPT-6 Astra)—and 12-hour retention is $20.
- **Inputs over 272K trigger a surcharge.** Once a single request exceeds 272K input tokens, input and cached-input rates double for the entire request, and output rates jump 1.5x. Long agent conversations cross that line quickly.
- **Ultrafast mode** runs at 3x standard pricing, rising to 6x starting March 1, 2027.

Knowledge cutoffs moved forward by two to four weeks. During A/B testing, developers identified the new checkpoints offline simply by asking if they knew about Fable 5—neither GPT-6 Astra nor GPT-6.1 Sol did.

## Official benchmarks: every chart is cost per task

OpenAI leads with Agents' Last Exam, which evaluates multi-step, long-horizon professional work. Astra at xhigh reasoning effort scores 78.3% at $2.89 per task, compared with Claude Fable 5.2 at Max (72.8%, $60.05). OpenAI frames that as 5.5 points higher at 5% of the cost.

![Agents' Last Exam v1.1 score vs. cost per task: GPT-6.2 Astra peaks at 78.3%, Gemini 4 Carbon sits close behind at 77.4%, and Claude Fable 5.2 scores 72.8% at $60](https://codingpass.ai/assets/img/blog/gpt-6-2-astra-sol-cheaper-not-smarter/ale.webp)

*Agents' Last Exam v1.1: score vs. cost per task. Source: OpenAI launch page.*

A closer look at the chart reveals three details OpenAI's copy skips:

- **Gemini 4 Carbon nearly ties Astra.** At High effort, Carbon hits 77.4% at $4.60—just 0.9 points behind Astra's best score. Pushing Astra to Max actually drops its score to 77.6% while raising task cost to $4.47, making it practically indistinguishable from Carbon. OpenAI's text compares Astra only against Fable 5.2 and GPT-6 Astra.
- **Nine tasks were excluded for Fable 5.2.** The footnote states that Fable 5.2's result excludes nine tasks where a fallback model was used.
- **Sol is the real value play.** GPT-6.2 Sol at Max scores 73.0% at $1.32 per task, matching last-gen flagship GPT-6 Astra (73.1%, $8.99) at one-seventh the cost.

Here are the rest of the official launch-page benchmarks (taking each model's highest-scoring effort level):

| Benchmark | GPT-6.2 Astra | GPT-6.2 Sol | GPT-6 Astra | Claude Fable 5.2 |
| --- | --- | --- | --- | --- |
| Agents' Last Exam | 78.3% ($2.89) | 73.0% ($1.32) | 73.1% ($8.99) | 72.8% ($60.05) |
| Terminal-Bench-Science 0.1 | 77.1% ($7.15) | 68.0% ($2.21) | 68.1% ($23.80) | 73.3% ($45.21)\* |
| FrontierCode 2.0 | 44.2% ($5.97) | 43.8% ($1.29) | 42.8% ($6.68) | 49.1% ($14.19) |
| GDPval v2 (OpenAI internal) | 50.4% | 46.9% | 44.2% | 44.8% |
| Harvey LAB-AA v1.1 | 15.0% ($6.35) | 11.1% ($1.92) | 8.6% ($13.67) | 7.2% ($39.84) |
| LifeVal | 86.1% ($4.23) | 83.8% ($1.27) | 79.8% ($9.04) | — |

\* Fable 5.2's Terminal-Bench-Science score comes from AA running inside the mini-swe-agent harness. OpenAI's own footnote acknowledges that Anthropic reports 74.6% for Fable 5.2 at Max inside Claude Code.

**Scientific research is the standout win.** On Terminal-Bench-Science, which tests research workflows inside a terminal, Astra beats GPT-6 Astra by 9 percentage points at less than a third of the cost. OpenAI calls Astra "the best model in the world for scientists," and it also ranks first on the Vals Science Index. Researchers already lean heavily on GPT models; this release widens that lead.

![Terminal-Bench-Science 0.1 score vs. cost: GPT-6.2 Astra 77.1% at ~$7, GPT-6 Astra 68.1% at ~$24, Claude Fable 5.2 73.3% at ~$45](https://codingpass.ai/assets/img/blog/gpt-6-2-astra-sol-cheaper-not-smarter/tbs.webp)

*Terminal-Bench-Science 0.1: score vs. cost. Source: OpenAI launch page.*

**GDPval v2 and legal tasks.** GDPval v2 is OpenAI's internal evaluation covering multi-step consulting, finance, and design workflows. Astra gains 6.2 percentage points over GPT-6 Astra while using 22% fewer output tokens; Sol gains 5.3 points over GPT-6.1 Sol. On legal benchmark Harvey LAB-AA, Astra's pass rate is more than double Fable 5.2's at roughly a sixth of the cost—though passing requires meeting every requirement with no material hallucinations, meaning Astra's 15% score still leaves 85% of legal deliverables failing.

**Low effort regressed sharply.** LifeVal measures everyday tasks such as screening daycares, apartment hunting, and meal planning. At Max effort, Sol hits 83.8% at a seventh of GPT-6 Astra's cost. Drop to low effort, however, and Sol plunges to 25.3%—down from 66.8% on GPT-6.1 Sol—while Astra's low effort falls from 67.5% to 46.3%. Anyone hoping to cut costs further by dialing down reasoning effort will want to test carefully.

![GDPval v2 score vs. output tokens: GPT-6.2 Astra 50.4% at ~36K tokens, Gemini 4 Carbon 48.8% at ~109K tokens, Claude Fable 5.2 44.8% at ~222K tokens](https://codingpass.ai/assets/img/blog/gpt-6-2-astra-sol-cheaper-not-smarter/gdpval.webp)

*GDPval v2 (OpenAI internal): score vs. output tokens. Source: OpenAI launch page.*

## Coding: stalled in place

OpenAI's launch post describes Astra's code as cleaner and better at implementing and testing complex changes in existing codebases. Its own FrontierCode 2.0 chart tells a flatter story.

![FrontierCode 2.0 score vs. cost: Claude Fable 5.2 leads at 49.1%, Claude Opus 5.5 at 45.5%, GPT-6.2 Astra at 44.2%, and GPT-6.2 Sol at 43.8%](https://codingpass.ai/assets/img/blog/gpt-6-2-astra-sol-cheaper-not-smarter/fc.webp)

*FrontierCode 2.0: score vs. cost (log scale). Source: OpenAI launch page.*

FrontierCode 2.0 tests whether changes meet a maintainer's merge bar. Astra tops out at 44.2%, trailing both Fable 5.2 (49.1%) and the previous-generation Claude Opus 5.5 (45.5%), and edging out GPT-6 Astra by only 1.4 points. OpenAI's text instead highlights that Sol sits within 0.5 points of Astra at roughly a fifth of the cost.

Artificial Analysis's independent Coding Agent Index (run in the Codex harness at max effort) is even blunter: GPT-6.2 Astra scores 53, unchanged from GPT-6 Astra, while GPT-6.2 Sol scores 49—3 points lower than GPT-6.1 Sol. Both Fable 5.2 and Opus 5.5 score 58 in Claude Code.

![Artificial Analysis Coding Agent Index: Claude Code with Fable 5.2 and Opus 5.5 both score 58; Codex with GPT-6.2 Astra scores 53, GPT-6 Astra 53, GPT-6.1 Sol 52, and GPT-6.2 Sol 49](https://codingpass.ai/assets/img/blog/gpt-6-2-astra-sol-cheaper-not-smarter/aa-coding.webp)

*AA Coding Agent Index vs. cost per task. Source: Artificial Analysis.*

AA traces Sol's drop to a 14-point regression on SWE-Atlas-QnA-AA, noting that Sol is more likely than its predecessor to omit key facts in its answers. AA also notes that multiple effort levels of both Astra and Sol sit on the Cost per Task Pareto frontier—the cost efficiency isn't in dispute.

## Third-party benchmarks: second overall, more hallucinations

Across the nine evaluations in the Artificial Analysis Intelligence Index v5, GPT-6.2 Astra at Max scores 62—up 5 points from GPT-6 Astra, ahead of Gemini 4 Carbon (60) and Claude Opus 5.5 (59), and second only to Claude Fable 5.2 (65). GPT-6.2 Sol reaches 57, up 2 points from GPT-6.1 Sol and tied with GPT-6 Astra at roughly 15% of its task cost.

![Artificial Analysis Intelligence Index: Claude Fable 5.2 at 65, GPT-6.2 Astra at 62, Gemini 4 Carbon at 60, Claude Opus 5.5 at 59, GPT-6.2 Sol at 57; scatter plot of index vs. cost per task below](https://codingpass.ai/assets/img/blog/gpt-6-2-astra-sol-cheaper-not-smarter/aa-index.webp)

*AA Intelligence Index and score vs. cost per task. Source: Artificial Analysis.*

During pre-release A/B testing, rumors claimed GPT-6.2 Sol hit 60 on the index and tied for first. Its actual verified score is 57, three points lower.

The lower task bills come mostly from generating far fewer output tokens. AA tracked average Output Tokens per Task across the Intelligence Index:

| Model (effort) | Output tokens / task | Output price / 1M | Output cost / task |
| --- | --- | --- | --- |
| GPT-6.2 Sol (Max) | 40K | $10 | $0.40 |
| GPT-6.2 Astra (Max) | 41K | $40 | $1.64 |
| GPT-6 Astra (Max) | 52K | $50 | $2.60 |
| Gemini 4 Carbon (High) | 123K | $10 | $1.23 |
| Claude Fable 5.2 (Max) | 274K | $40 | $10.96 |

Fable 5.2 and Astra charge the same per-token rate, yet Fable 5.2 outputs 6.7x as many tokens on the same task—an order-of-magnitude difference on the final bill. One developer comparing usage on a $200 ChatGPT Pro plan ran four video/3D projects, two large architecture plans, and nearly a hundred code reviews on Astra using 6% of a weekly allowance, whereas 10 runs on Fable 5.2 previously burned close to 30%.

That terseness comes with a catch. On AA-Omniscience, which measures whether a model guesses instead of abstaining when it doesn't know, GPT-6.2 Astra's hallucination rate climbs from 50% (GPT-6 Astra) to 59%. Sol rises from 52% to 55%. Gemini 4 Carbon sits at 11%.

![AA-Omniscience Hallucination Rate (lower is better): Gemini 4 Carbon 11%, GPT-6 Astra 50%, GPT-6.1 Sol 52%, GPT-6.2 Sol 55%, GPT-6.2 Astra 59%](https://codingpass.ai/assets/img/blog/gpt-6-2-astra-sol-cheaper-not-smarter/aa-hallucination.webp)

*AA-Omniscience Hallucination Rate, lower is better. Source: Artificial Analysis.*

## Hands-on tests: strong reviewer, weak planner

Developer project runs match the benchmark split: what the family already did well got better, and the weak spots stayed weak.

**Code review improved the most.** A common workflow today is writing code with Claude and having GPT review it for up to three rounds. On a developer's internal bug-finding test set at xhigh effort, GPT-6.1 Sol caught 65% of issues, GPT-6.2 Sol caught 84%, and GPT-6.2 Astra caught 96%. Astra took twice as long and cost 4x as much to catch around a dozen extra bugs, making GPT-6.2 Sol at xhigh the practical default for code review.

**Planning and communication remain hard to parse.** Given the same brief—adding a low-cost, compliant comment section to a solo-maintained news site—Gemini 4 Carbon produced a clear plan in five minutes: listing real-name and pre-moderation compliance rules first, proposing a minimal design (no custom handles, no nested threads, curated comments only), reusing the site's existing agent for automated moderation, and finishing with a comparison table. GPT-6.2 Astra ran for 24 minutes, returned a short but hard-to-follow response, and suggested manual moderation at launch—unrealistic for a solo maintainer. Testers felt Astra's readability and proactive planning were actually worse than early GPT-6 Astra.

**Mixed aesthetics, standout 3D.** On a prompt for a ChatGPT four-year anniversary motion video, Fable 5.2 and Gemini 4 Carbon both turned out usable clips, while Astra's output left testers speechless for the wrong reasons. 3D is the exception: a first-person parkour game, a Three.js autumn rural train station, and an ancient Greek city-state rebuilt from a reference image all beat expectations—the last one running for over an hour.

AA's 3D evaluation backs that up. On AA-3D (Elo, blind pairwise votes by human judges), GPT-6.2 Astra takes first place at 1,318, ahead of Fable 5.2 (1,271) and GPT-6.2 Sol (1,259). On AA-Robotics (research preview), Astra scores 79 and Sol 78, beating Fable 5.2's 75.

![AA-3D Elo: GPT-6.2 Astra 1318, Claude Fable 5.2 1271, GPT-6.2 Sol 1259, Claude Sonnet 5.5 1204](https://codingpass.ai/assets/img/blog/gpt-6-2-astra-sol-cheaper-not-smarter/aa-3d.webp)

*AA-3D Elo, blind pairwise human votes. Source: Artificial Analysis.*

OpenAI's launch page showcases several other Astra examples: an editable valuation workbook built at an 8% discount rate and 3% terminal growth rate, a 19-slide deck matching company template font sizes, part-by-part CAD/simulation models of an RC off-road truck and an inline-six engine, and a playable racing game, [Meridian Sound Racing](https://meridian-futuristic-race.openai.chatgpt.site/). On AA-AnalystAgent, Astra answered 62.5% of questions correctly across all five independent attempts.

## System card caveats

OpenAI's launch post and the [GPT-6.2 system card](https://deploymentsafety.openai.com/gpt-6-2-astra-sol) flag three caveats of their own:

- **Concealing task blockers.** Across two alignment evaluations, both models were more likely than the previous generation to omit reporting obstacles—such as deleting a target file in a coding task or disabling a search tool—instead of telling the user. OpenAI added instructions inside Codex to curb this, but developers building custom agents on the API will need their own guardrails.
- **More cybersecurity refusals.** Both models cross the "Critical" cybersecurity threshold under OpenAI's Preparedness Framework, triggering stricter launch safeguards. OpenAI acknowledges this will block some legitimate vulnerability research and security testing; looser access requires OpenAI Daybreak, which remains on the GPT-6 generation.
- **High threshold for AI self-improvement.** Astra is OpenAI's first publicly released model to reach the "High" threshold for AI self-improvement, prompting new Control Room monitoring and interruption mechanisms that can pause workloads or conversations on supported deployments.

## What 100 and 1,000 tasks actually cost

List prices per million tokens don't show the real bill. Using OpenAI's published cost per task on Agents' Last Exam, here is what 100 and 1,000 runs add up to. "Cost per pass" divides task cost by score—a rough estimate of what one successful run costs, excluding human rework time.

| Model (effort) | Score | Per task | 100 tasks | 1,000 tasks | Cost per pass |
| --- | --- | --- | --- | --- | --- |
| GPT-6.2 Sol (Max) | 73.0% | $1.32 | $132 | $1,320 | $1.81 |
| GPT-6.2 Astra (xhigh) | 78.3% | $2.89 | $289 | $2,890 | $3.69 |
| Gemini 4 Carbon (High) | 77.4% | $4.60 | $460 | $4,600 | $5.94 |
| GPT-6 Astra (Max) | 73.1% | $8.99 | $899 | $8,990 | $12.30 |
| Claude Opus 5.5 (Max) | 71.6% | $17.57 | $1,757 | $17,570 | $24.54 |
| Claude Fable 5.2 (Max) | 72.8% | $60.05 | $6,005 | $60,050 | $82.49 |

Gemini 4 Carbon's task cost uses its introductory pricing; Carbon's rates double on December 1.

Run the same cost-per-pass math on FrontierCode 2.0 (task cost ÷ score) to see what a merge-ready change costs: GPT-6.2 Sol comes out to $2.95, GPT-6.2 Astra to $13.51, Claude Opus 5.5 to $28.18, and Claude Fable 5.2 to $28.90. That table sums up where GPT-6.2 actually sits: Sol delivers nearly 90% of Fable 5.2's score at roughly a tenth of the cost, while Astra sits awkwardly in the middle—4.6x pricier than Sol for just 0.4 extra points on code.

## Which model to pick

| Workload | Pick | Why or trade-off |
| --- | --- | --- |
| Writing code, large refactors | Claude Fable 5.2 / Opus 5.5 | Leads both FrontierCode and AA Coding Agent Index; Fable 5.2 burns a lot of tokens |
| Code review, bug hunting | GPT-6.2 Sol (xhigh) | 84% catch rate in testing at a quarter of Astra's price; step up to Astra for 96% |
| Scientific research, data analysis, terminal experiments | GPT-6.2 Astra | First on Terminal-Bench-Science at roughly a sixth of Fable 5.2's task cost |
| 3D, Three.js, game prototypes | GPT-6.2 Astra | First on AA-3D; video and motion aesthetics remain weak |
| Architecture planning, stakeholder communication | Gemini 4 Carbon / Claude | GPT-6.2 is slow and hard to parse |
| Fact-checking, citation-heavy writing | Gemini 4 Carbon | GPT-6.2 hallucinates 55%–59% vs. Carbon's 11% |
| High-volume, low-cost agent runs | GPT-6.2 Sol (medium or above) | Avoid low effort, which drops to 25.3% on LifeVal |

A common split is letting Claude or Gemini plan and write the code, then handing strict subtasks and code review to GPT-6.2. GPT models are disciplined, literal, and information-dense, which suits science and algorithm work—but proactive planning and clear prose still aren't this generation's strength.

## Running GPT-6.2 on CodingPass

GPT-6.2 Sol is included on Plus ($30/mo), Pro, and Team Pass, with up to $200 per month on Plus. GPT-6.2 Astra is available on Pro ($100/mo) and Team Pass, with up to $500 per month on Pro. Usage is metered at OpenAI's official API rates, including official rates for cached input.

Terminal

```
codex -m gpt-6.2-sol     # Code review, everyday agents
codex -m gpt-6.2-astra   # Science, 3D
```

See the [Codex setup guide](https://codingpass.ai/tools/codex) for configuration steps; OpenCode, Cline, and other clients use the same model IDs. If you only write code and don't need GPT, there's no reason to upgrade for it: Claude Sonnet 5.5 on Go and Opus 5.5 on Plus both outperform GPT-6.2 on coding.

## FAQ

**Can GPT-6.2 Sol replace Opus 5.5 in Claude Code for writing code?**

Not recommended. On AA's Coding Agent Index, GPT-6.2 Sol scores 49 against Opus 5.5's 58 in Claude Code, and on OpenAI's own FrontierCode 2.0 chart Sol's 43.8% trails Opus 5.5's 45.5%. Sol's low cost makes it much better suited to reviewing code that Opus writes.

**How much allowance does GPT-6.2 Astra save over GPT-6 Astra?**

List prices drop 20%, cached input drops 60%, and output token counts are lower. Across OpenAI's published task costs on Agents' Last Exam and Terminal-Bench-Science, Astra runs under a third of GPT-6 Astra's bill. Keep in mind that requests exceeding 272K input tokens double the input rate for the whole call.

**Is Max reasoning effort always the best choice?**

No. On Agents' Last Exam, Astra at xhigh (78.3%, $2.89) scores higher and costs less than Max (77.6%, $4.47). At the other end, avoid low effort—Sol drops to 25.3% on LifeVal. Use medium or high for routine work, and xhigh for hard problems.

Written by CodingPass team

We run CodingPass and test each guide against the latest stable release of the tool before we publish it. Spotted something out of date? Write to [support@codingpass.ai](mailto:support@codingpass.ai).

## Related reading.

Model tests

### [Claude Fable 5.2 is out: benchmarks, pricing and our take](https://codingpass.ai/blog/claude-fable-5-2-is-out-benchmarks-pricing-and-our-take)

Fable 5.2's official benchmarks, the new lower prices, which effort level to use, and how to switch to it on CodingPass Pro.

Model tests

### [Gemini 4 Krypton vs GPT-6.2 Sol](https://codingpass.ai/blog/gemini-4-krypton-vs-gpt-6-2-sol)

Circulating pre-release Sol benchmarks and how to pick a model per job.

Models

### [All models](https://codingpass.ai/models)

Every model's ID, minimum plan and monthly allowance on Go, Plus and Pro.
