跪拜 Guibai
← All articles
ChatGPT · GPT

GPT-6.2 Hits the API: Sol Cuts Costs, But a Stealthy Cache-Busting Change Sneaks In

By 一缕82年的清风 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Anyone running OpenAI on production workloads will pay more or lose performance if they miss the cache-breaking instructions change, which turns cheap cached tokens back into full-price calls. The Sol-vs-Astra benchmark gap also makes a concrete cost case for switching most coding traffic to the mid-tier model before the 3x ultrafast billing window ends.

Summary

OpenAI released GPT-6.2 on November 17 with two models: the flagship Astra and the mid-tier Sol. Both share a roughly 1.05 million token context window and a June 2026 knowledge cutoff, but pricing and positioning split cleanly. Astra drops its input/output rates about 20% from the original GPT-6 Astra, while Sol keeps the familiar $2/$10 per million tokens and produces tighter output. The real headline is Sol: its FrontierCode 2.0 score sits 0.4 points behind Astra (43.8% vs 44.2%) at a quarter of the price, and its GPQA Diamond trail is just one point. Sol also supports an Ultrafast mode that pushes throughput to about 8x standard, though the API bills it at 3x until February 2027 and 6x afterward.

Community reaction splits by model. Codex users running Sol in ultrafast mode report good results, while Astra draws criticism for over-engineering, layering abstractions, and asking unnecessary confirmations before editing code. Benchmarks show Astra pulling ahead mainly on scientific reasoning (Terminal-Bench-Science 0.1: 77.1% vs 68.0%), which suggests most teams should route ordinary coding through Sol and reserve Astra for research-heavy work.

Migrating to 6.2 involves three API gotchas. Top-level instructions are now concatenated after the input array, which changes the request prefix and invalidates cached prefixes; the fix is to place reusable system rules in the first role: "developer" message. Passing reasoning.effort: "none" now errors out; valid values are low through max. And editing conversation history or switching models mid-run clears the reasoning cache. The article includes a 12-line Python template that demonstrates the correct caching-safe call, along with routing advice that keeps Claude, Gemini, and DeepSeek models in the gateway for resilience.

Takeaways
— GPT-6.2 Sol nearly matches Astra on code benchmarks: FrontierCode 2.0 43.8% vs 44.2%, GPQA Diamond 96.0% vs 97.0%, while Sol costs $2/$10 per million tokens versus Astra's $8/$40.
— Astra's real edge is only in scientific terminal tasks: Terminal-Bench-Science 0.1 scores 77.1% for Astra versus 68.0% for Sol.
— Top-level instructions in the Responses API are now appended after the input array, so any shared system rules must go into the first role: "developer" entry to preserve prompt caching.
— reasoning.effort: "none" is no longer accepted in GPT-6.2; valid values are low, medium, high, xhigh, and max.
— Ultrafast mode delivers about 8x throughput but is billed at 3x through February 2027 and 6x from March 1, 2027.
— Batch requests get half price, and Background windows (2h/8h) get a 30-40% discount.
— The included 12-line Python example shows the correct pattern: developer message first, user message second, one-off instructions in the top-level parameter.
Conclusions

The near-tie on coding benchmarks makes Sol the obvious default for most engineering teams, but the 3x-to-6x ultrafast billing scheme still rewards teams that batch or defer heavy work until after the promotional window.

OpenAI's decision to silently change where top-level instructions sit is a reminder that API upgrades carry hidden runtime costs that do not show up in benchmark tables.

Community rants about Astra's over-engineering suggest a stability-vs-verbosity tradeoff: the flagship may be tuned for research scenarios where extra clarification matters, which irritates developers doing routine code edits.

Keeping multiple model vendors in the gateway, as the article recommends, is a practical hedge against exactly this kind of mid-version API drift.

Concepts & terms
Prompt caching
A mechanism where the API reuses computed results for a repeated request prefix, charging a lower cache price instead of the full input price. It breaks whenever the prefix changes, such as adding instructions after the input.
Ultrafast mode
An optional GPT-6.2 mode that raises throughput to roughly 8x standard, but at a higher token price that changes from 3x to 6x after a promotional period.
Terminal-Bench-Science
A benchmark that evaluates AI agents on scientific research and laboratory tasks, distinguishing specialized performance from general coding ability.
From the discussion
Featured comments
用户368345607470

As cited in the article, Sol and Astra differ by only 1 point on GPQA, 0.4 points on FrontierCode, but 9.1 points on terminal science benchmarks. Model selection should be based on task type rather than just price. Ultrafast claims 8x throughput, but as the article states, until the end of February 2027 it is billed at 3x. Whether the speed benefit justifies the 3x price depends on real measurements of queue time, time-to-first-token, and latency distribution.

See top comments, translated →
Source: juejin.cn ↗ Google Translate ↗ Backup ↗