GPT-6.1 Sol vs GPT-6 Astra for CV tailoring

What changed when OpenAI launched GPT-6.1 Sol?

OpenAI launched GPT-6.1 Sol on 29 September 2026 at its DevDay event, and the pitch is blunt: near-Astra intelligence for a fifth of the price. Standard API rates are $2 per million input tokens and $10 per million output, against $10 and $50 for GPT-6 Astra. Cached input drops to $0.10 per million — a tenth of Astra's $1. The model carries a context window of roughly one million tokens, supports up to 128,000 output tokens, and has an April 2026 knowledge cutoff. At launch it reached Plus, Pro, Business, Enterprise and Edu users inside ChatGPT Work and Codex rather than the regular chat window, plus the API. No GPT-6.1 Astra shipped alongside it.

The benchmark numbers back most of the marketing. On DeepSWE v1.1, which runs long software-engineering tasks in real codebases, OpenAI says Sol matches Astra at about a fifth of the cost and beats GPT-6 Sol's best result by 6.4 percentage points. On OSWorld 2.0 for computer use it lands 2.1 points behind Astra at roughly a seventh of the cost. On GDP.pdf — professional documents thick with tables, charts and fine print — Sol scored 32.0% at high effort versus Astra's 31.0%, at $0.35 per task against $1.79. Astra still posts the top score on Terminal-Bench Science at 68.1%, and OpenAI still points to it for the hardest research work.

None of that is why this matters to you, though. It matters because the price of the model you paste job ads into just stopped being a reason to cut corners. A tailored CV is a small job by frontier-model standards: your CV plus one job ad is a few thousand tokens, and the rewrite is a few thousand more. Whichever model you use, the per-application cost is pennies. So the real question stops being "can I afford the good one" and becomes "which of these three CV jobs actually needs the expensive model" — rewriting bullets, pulling keywords out of a job ad, or tailoring an entire document end to end.

Which model should rewrite your CV bullet points?

For rewriting bullet points, GPT-6.1 Sol is enough, and paying five times more for Astra to do it is the most wasteful line in an AI-assisted job search. Rewriting a bullet isn't a reasoning problem. You already know what you did; the model is compressing it into a line a recruiter can scan in two seconds. Both models handle that at medium reasoning effort without breaking a sweat, and the difference in output quality is smaller than the difference between a good prompt and a lazy one. Sol has five reasoning levels — low, medium, high, xhigh and max — and bullets rarely need more than medium.

What does move the needle is what you feed it. "Make my CV bullets stronger" produces adjectives. "Here are the raw facts: I ran a team of four, we cut invoice processing from nine days to four over eight months, and the queue dropped from 400 items to under 60 — write three bullets for this job ad" produces something a hiring manager believes. Give the model the ad, the numbers, the tools you actually used and the constraint you worked under. The model's job is phrasing and keyword alignment, not memory. It cannot recover a metric you never told it, and it will happily paper over the gap with vague language.

One practical trick: ask for three versions of each bullet at different lengths, then pick. Sol's cheap output pricing means generating variants costs almost nothing, and comparing them side by side exposes the weak claim fast — the version that sounds strongest but says least is usually the one you were about to keep. If you'd rather skip prompt wrangling entirely, an AI CV builder that works by chat does the same loop with structure built in: paste an old CV or describe the experience, refine by conversation, then export to PDF or Word in one of six templates.

Does the cheaper model read a job ad as accurately?

Yes — for pulling requirements and keywords out of a job ad, GPT-6.1 Sol is at least as reliable as Astra, and it's the cheaper model's strongest use case. Extraction is reading comprehension over a short document, which is exactly where Sol closed the gap. OpenAI reports that at low reasoning effort the share of responses containing at least one factual error fell from 11.4% with GPT-6 Sol to 7.7% — about a 32% reduction — and that across reasoning settings the error rate stays within 1.9 percentage points of Astra. On document-heavy professional tasks it edged Astra at high effort while costing roughly a fifth as much per task.

The prompt that works is boring and specific. Paste the full ad and ask for three lists: hard requirements stated as must-haves, the exact tool and technology names as written, and the soft signals repeated more than once. Ask it to quote the ad's own wording rather than paraphrasing, because the wording is the point — an applicant tracking system matching "stakeholder management" won't credit you for "worked with lots of teams". Then ask which of those items your current CV never mentions. That last question is the one most people skip, and it's the one that produces an actual to-do list.

Two honest caveats. First, absolute scores on strict professional-document benchmarks still sit in the low thirties for every frontier model, meaning all of them fail the majority of hard document tasks when judged on every criterion. A model reading a job ad well is not the same as a model knowing how a parser will read your file. Second, if you paste a LinkedIn or Indeed link instead of the text, fetching usually fails — copy the ad body in as plain text. Miss that and you'll get confident analysis of a page the model never actually saw.

Where does GPT-6.1 Sol flatten your impact numbers?

The cheaper model's weak spot isn't vocabulary — it's holding a long, messy career in its head while deciding what to cut. Hand Sol a fifteen-year CV, four target roles and an instruction to compress it to two pages, and it tends to average everything: strong outcomes get the same weight as routine duties, sharp numbers get rounded, and the story of why you moved between roles goes missing. The benchmark pattern matches. Astra still leads on the hardest multi-step reasoning tasks, topping Terminal-Bench Science at 68.1% while costing about $23.80 per task against $5.47 for Sol. You pay four times more for judgement, not for prose.

Watch for three specific failure modes in the output. Numbers drift — "cut onboarding from 14 days to 6" becomes "reduced onboarding time by over 50%", which is weaker and harder to defend in an interview. Scope inflates — "supported a migration" becomes "led the migration", and that's the kind of line that collapses under one follow-up question. And hedging creeps in, where a concrete result turns into "contributed to significant improvements". Every one of these makes your CV read like a generated document, which is now a live screening concern rather than a theoretical one.

So run a two-model split if you have access to both, and don't overthink it. Use Astra once, at the start, for the structural call: which roles get real estate, what order the sections go in, what gets cut to a single line. Then switch to Sol for every rewrite, every keyword pass and every per-application tailor. If you only have Sol, break the work into smaller pieces instead — one role at a time, one job ad at a time — because Sol's drift shows up under load, not in short tasks.

How do you check what the model produced before applying?

Check the file, not the chat. A model's own verdict on its work is the least reliable signal available, because it's grading the text it just wrote against criteria it inferred — and it has no visibility into how a parser will extract your PDF. The three things worth verifying independently are whether the file's text comes out in the right order, whether the ad's actual keywords made it in, and whether every claim traces back to something true. Do that in a separate step, with a separate tool, on the exported document rather than on the answer in the chat window.

A free CV analysis is the fastest way to do it. Upload the PDF, Word file or even a photo — Hebrew OCR included — and you get an overall score out of 100 rated Weak, Decent, Strong or Excellent, five category scores covering experience, tech stack, impact and ownership, clarity and structure, and ATS compatibility, plus your strengths and a visual layout analysis. Paste the job description text alongside it and the analysis names the skills you're missing and tailors its suggestions to that ad. Every analysis also generates ATS-friendly rewritten CVs across six templates that mirror the ad's keywords.

If you want the diagnosis rather than just the score, the Full Analysis at $3.99 adds the fix list, the blunt "why you might be rejected" read, blind spots you can't see from inside your own CV, a hiring probability expressed as interviews per ten applications against the market, and a salary estimate. That's the loop worth building: model drafts, tool scores, you fix, then apply. It's also the honest answer to the model-choice question people obsess over — and if you're weighing assistants more broadly, how ChatGPT and Claude each handle CV writing matters less than whether anyone checked the file you sent.

Frequently asked questions