QWE AI Academy Should you still pay Astra prices now that GPT 6.1 Sol is live?
Key takeaway: Switch most coding agents, PDF/document jobs, and multi-step work to GPT-6.1 Sol. Sticker rates as of late September 2026: $2 input / $10 output per million tokens versus Astra’s $10 / $50, with cached input at $0.10. Keep Astra when a couple of points on brutal computer-use or science runs actually change the outcome.
It shipped at DevDay – about seven days after GPT-6 Sol – and teams tore out the old model string the same afternoon. I re-ran a messy multi-file refactor and a PDF-heavy brief both ways. Below is what changed for real use, not the launch slide.
What just dropped (the 60-second version)
$2 / $10 standard tokens. $0.10 cache reads – half what GPT-6 Sol charged for the same hit, and 95% off Sol’s own standard input. That’s the headline from the OpenAI launch post: near-Astra agentic coding, computer use, and professional work at about one-fifth Astra’s standard input/output prices.
Model ID: gpt-6.1-sol. Context 1.05M (922K max input, 128K max output). Knowledge cutoff April 30, 2026. Surfaces as of launch: ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu. Regular Chat – not yet. API string is the same ID.
Artificial Analysis puts it one Intelligence Index point under Astra, with max-effort tasks around $0.72 vs Astra’s $3.26. Discord summary I kept hearing: “good enough that Astra feels like a luxury tax.”
Method A vs Method B: stay on Astra or move?
Same two workloads. Two paths.
| Path | When it wins | What it costs you |
|---|---|---|
| Method A – GPT-6 Astra | Hard science-style prompts and top-end computer use where a point or two matter | $10 / $50 per 1M tokens ($1 cached input) |
| Method B – GPT-6.1 Sol | Coding agents, PDF/document work, most business flows, budget computer-use agents | $2 / $10; cache $0.10; OpenAI’s DeepSWE v1.1 chart: match to Astra at ~1/5 cost, +6.4 pp over best GPT-6 Sol |
OSWorld 2.0 offline (OpenAI numbers): Sol within 2.1 points of Astra at roughly one-seventh the cost per task, and +7 pp over GPT-6 Sol at max. My refactor? Method B. Long simulation-style science prompt? Method A still felt sharper. Default for ~80% of my week: Sol.
How to actually use GPT 6.1 Sol
Model picker lit up. This is the path that worked.
In ChatGPT Work or Codex
- Open Work (or Codex desktop / CLI / IDE extension).
- Model control under the composer (or Advanced / Power settings).
- Select GPT-6.1 Sol. Enterprise/Edu: confirm an admin enabled the model for the workspace if it doesn’t appear.
- Reasoning effort: start high for coding and documents; max for heavy computer-use or science. API default is medium – already solid.
codex --model gpt-6.1-sol
# one-shot
codex exec -m gpt-6.1-sol "Review the current changes"In the API
Tools need the Responses API – model docs are blunt about that. Chat Completions is fine only when you’re not tool-calling.
from openai import OpenAI
client = OpenAI()
resp = client.responses.create(
model="gpt-6.1-sol",
reasoning={"effort": "high"},
input="Find the root cause of the flaky checkout test and propose the smallest fix."
)
print(resp.output_text)Supported reasoning.effort values: low, medium (default), high, xhigh, max. none and minimal are gone – old GPT-6 Sol agents that leaned on none for cheap tool loops will 400 until you rewrite that field.
Pro tip: Pin coding agents to high first. OpenAI’s DeepSWE chart peaks at high (75.2% at ~$0.65 in the launch figures via community write-ups); xhigh/max slide to 71.9% and spend more. Max is not a free upgrade.
Cache is where the math flips. Resend the same system prompt + repo prefix and you pay $0.10 cached input instead of $2. Stable prefixes matter more than clever temperature knobs. Cache writes list at $2.50 per 1M if you’re tracking full cost.
One-line prod migration for a typical agent loop: swap the model id, audit effort, keep prefixes cacheable. Codex long-context runs and Work PDF Q&A are the boring wins.
Edge cases that bite after you switch
The catch is the launch threads stay quiet on these.
- The 272K cliff. Cross 272K input tokens and the entire request bills at 2× input/cache and 1.5× output – not only the overflow. One fat agent turn doubles the line item. Stay under on purpose, or accept the step-up.
- Chatty outputs vs old Sol. Artificial Analysis saw ~10-30% more output tokens across effort levels. Cache-heavy agents still win; one-off chatty runs give some of the “cheap” feeling back.
- Tools = Responses only. Completions without tools: fine. Function-calling loops: Responses. Migrate before you flip the default.
- Fast / Batch / Flex. Fast is 2× Standard. Batch and Flex are 50% off Standard. Don’t mix those multipliers up when you forecast.
- Not in Chat (yet). Everyday Chat still shows older options. Work and Codex are the live surfaces.
Ultrafast is listed as “coming days” with up to 8× faster token generation in Codex. Budget on Standard until that ships.
There’s a strange calm after the switch: answers feel familiar, the invoice shrinks, and Tuesday’s PR review no longer looks like an Astra-shaped hole in the budget.
FAQ
Is GPT 6.1 Sol really near-Astra for coding?
On OpenAI’s DeepSWE v1.1 chart, yes – match at ~1/5 cost. Still run your own suite; launch benches aren’t your traffic.
Which reasoning effort should I pick first?
High for coding and document/PDF work. Max when computer-use or science prompts need every point OSWorld-style evals reward. Medium if you care about latency first. I had a multi-file bug hunt finish cleaner at high than at max – max burned tokens restating the plan.
Do I need to change my GPT-6 Sol API code?
Three checks only:
- Model string →
gpt-6.1-sol - Drop
reasoning.effort: "none"/minimal - Tool loops → Responses API if they still sit on Chat Completions
Sticker $2/$10 stays; cache reads at $0.10 reward stable prefixes. Batch remains half price when you can wait.
Open Work or Codex, pick GPT-6.1 Sol at high, rerun one task you usually burn Astra on. Quality plus the usage line – that single A/B sets the week’s default.