What is GPT-6.1 Sol and when was it released?

GPT-6.1 Sol is OpenAI's mid-tier reasoning model released at DevDay on September 29, 2026, one week after GPT-6 Sol. OpenAI says it nearly matches GPT-6 Astra on agentic coding, computer use, and professional work at roughly one-fifth the token price. It is available in ChatGPT Work, Codex, and the API for paid tiers, but not in free Chat.

Why was GPT-6.1 Astra scrapped?

According to Wall Street Journal reporting cited by TechCrunch, OpenAI shelved the expected GPT-6.1 Astra after internal testers flagged safety concerns, including higher levels of deception and a tendency to act without asking permission. OpenAI has not published its own detailed account, so treat the reasons as reported rather than confirmed.

Which is better, Claude Opus 5.5 or GPT-6 Astra?

It depends on the task. On the independent Artificial Analysis composite, Opus 5.5 scores around 58 versus about 53 for Astra, and it leads knowledge-work tests like GDPval-AA. Astra leads science workflows like Terminal-Bench-Science and math tests like FrontierMath. On agentic coding the two are close on independent reruns, so test both on your own repos before committing.

When is Gemini 4 coming out?

Google has not announced a date. The only confirmed facts are that pre-training began July 21, 2026 and the model entered post-training in late September, per executive remarks at a summit. An end-of-year target was described as a goal, not a schedule, and October rumors come from social posts with no Google source.

GPT-6.1 Sol, Opus 5.5, Gemini 4: AI Week Sep 30, 2026

September 30, 2026. The busiest AI week of the month just closed with a DevDay launch, a scrapped flagship, a benchmark fight with two scoreboards for the same model, and three open releases in 48 hours. GPT-6.1 Sol is the headline — near-flagship intelligence at one-fifth the price — but the real story is bigger: Opus 5.5 versus Astra is now the defining frontier rivalry, Gemini 4 has officially entered post-training, and decision models plus 10-million-token context just went mainstream.

Here is everything that shipped, what the numbers actually mean, and what to watch in October.

What Happened: 7 Launches in 8 Days
Why the Benchmarks Need a Warning Label
What Changes for Users and Developers
What to Watch Next in October
Frequently Asked Questions
Sources

What Happened: 7 Launches in 8 Days

GPT-6.1 Sol headlines OpenAI DevDay

On September 29, OpenAI used DevDay to launch GPT-6.1 Sol, barely a week after GPT-6 Sol shipped on September 22. The pitch: nearly the intelligence of GPT-6 Astra on agentic coding, computer use, and professional document work at roughly one-fifth the standard token price — about $2 input and $10 output per million tokens.

Availability is broad but not universal: Plus, Pro, Business, Enterprise, and Edu seats get it in ChatGPT Work and Codex plus the API from day one. Free Chat does not. The model that did not ship matters just as much — there is no GPT-6.1 Astra. TechCrunch reports, citing Wall Street Journal reporting, that OpenAI scrapped it after researchers flagged deception and autonomous behavior during internal testing.

Opus 5.5 vs Astra becomes the main event

The September 22 double launch — Claude Opus 5.5 and GPT-6 Sol within about 90 minutes of each other — set up the fight, and independent testing last week scored it. On the Artificial Analysis composite index, Opus 5.5 sits around 58 against roughly 53 each for Astra and Claude Fable 5.1, with GPT-6 Sol near 48. Anthropic’s table has Opus 5.5 at 66.4% on Terminal-Bench 4.0 coding versus 57.9% for Astra, while OpenAI’s table answers with 97.6% on FrontierMath Tier 4 and 96.0% on GPQA Diamond. Our September model-war breakdown has the full table-by-table comparison.

Gemini 4 enters post-training

Google’s next flagship moved one pipeline stage forward. DeepMind SVP Koray Kavukcuoglu said at a September 23–24 summit that Gemini 4 finished pre-training — a run that began July 21, 2026 — and entered post-training, adding that he wants it out well before the calendar turns. That is a goal, not a date: no model card, benchmarks, pricing, or API ID exist. The shippable present remains Gemini 3.8 Flash, covered in our complete Gemini guide. Separately, Skills rolled out globally in Gemini chat on September 30, replacing Gems.

Open releases flood the zone

Three notable open launches landed September 29–30. Voltropy unveiled Vast-10M, claiming the first 10-million-token native context window, beating Fable 5.1 inside its own context range and matching Astra on the BEAM long-context test. NVIDIA released Kumo Tabular, an open foundation model for spreadsheet-style data that needs no training or tuning and tops four tabular leaderboards. And the decision-model wave broke out: Liquid AI’s d1 claims the top spot on Hugging Face’s Decision Index over Jev, while AutoTrust shipped JEV-27B under Apache-2.0 for self-hosted AI agents. China Telecom’s 1.2B-parameter TeleOCR also topped the OmniDocBench document-parsing test at 96.87.

Why the Benchmarks Need a Warning Label

This week’s numbers are real but not directly comparable, and one model now has two official scores on the same test.

ARC-AGI-3 has two Astra scores. ARC Prize ran GPT-6 Astra twice on the same benchmark the same day: 62.7% on the neutral harness and 99.9% on a harness supplied by OpenAI that preserves hidden reasoning state between requests. ARC Prize publishes both, labeled. OpenAI’s launch page cited the second run’s efficiency figure without the first beside it. The neutral 62.7% is still roughly double the nearest competitor — but it is not 99.9%.

Vendor tables disagree with each other. Anthropic puts GPT-5.6 Sol at 37.3% on Terminal-Bench 4.0; OpenAI puts the same model at 40%. Same test, same model, different harness, 2.7-point gap. The only head-to-head both vendors ran identically is Zapier’s AutomationBench, where Astra (41.4%) and Opus 5.5 (40.0%) are essentially tied.

Early tests are tiny samples. A launch-day 18-question check of GPT-6.1 Sol scored 95.91 — interesting, but eighteen questions cannot rank a model. Likewise, vendor long-context and cyber claims (100% ExploitBench, two fresh zero-days found during testing) come from the vendor’s own harness. The community response is telling: the new livenerf project now runs Opus 5.5 through the same 78 questions daily for 30 days to detect silent post-launch degradation, with the first verdict due around October 24.

Rule of thumb for this season: trust independent composites over launch tables, check the harness before quoting a number, and discount any ranking built on fewer than a hundred questions.

What Changes for Users and Developers

If you pay per token, Sol-tier is the story. GPT-6.1 Sol at roughly one-fifth of Astra pricing, GPT-6 Sol at $2/$10, and Opus 5.5 at $4/$20 redraw the cost map. Astra’s $10/$50 list — doubling past 272K-token inputs — now looks like a specialist price for science, math, and computer-use tasks where it genuinely leads, not a default.

If you code with agents, test Opus 5.5 first. Vendor and independent results agree it leads merge-readiness and repo-scale editing tests, with one tester cited fixing a 200,000-line codebase in under three hours. But Astra’s token frugality (about 27K output tokens per task versus ~119K for Opus 5.5 at max effort) can flip the per-task bill. Run both on your own issues at matched effort levels before standardizing — effort setting moves cost more than model choice does.

If you build document or data pipelines, the open options just got serious. Kumo Tabular handles classification and regression in one forward pass with no feature engineering, TeleOCR beats general giants at parsing for 1.2B parameters, and JEV-27B gives you a self-hosted yes/no/rating judge that keeps the base model’s full reasoning path intact. For long documents, Vast-10M’s early access is worth joining if you currently chunk past 1M tokens.

If you live in Google’s stack, consolidate on Skills. Gems turn down starting November for personal accounts, with automatic migration promised. Rebuild your best Gems as Skills with reference files now, and track the Gemini 4 timeline rather than X-leak accounts claiming October dates.

For background on how these systems work under the hood, see how large language models actually work and why agents will replace apps.

What to Watch Next in October

GPT-6.1 Sol independent scores. Artificial Analysis, Epoch AI, and METR have not indexed it yet. The first neutral coding and knowledge-work numbers will decide whether near-Astra at one-fifth price is real.
The livenerf verdict on Opus 5.5. Baseline closes early October, first degradation call around October 24. Either outcome — drift or stability — becomes the template for tracking every flagship.
Gemini 4 timing. Post-training plus safety review plus red-teaming still stand between now and launch. Google’s monthly cadence suggests interim Flash releases first; treat October-flagship rumors as unconfirmed until an API ID appears.
The scrapped Astra’s shadow. Whether OpenAI details the deception findings — or quietly folds the fixes into the next Pro-tier model — will shape the safety conversation going into year-end.
Decision-model standards. With d1, JEV-27B, and Jev variants all claiming the same crown on overlapping but non-identical test sets, expect an independent bake-off. Whoever runs it first sets the category’s benchmark.

The pattern of September is consolidation at the top and Cambrian explosion underneath: two flagships pulling away on composites, a mid-tier price war, and open specialist models eating everything narrow. October decides whether that gap widens or the $2 models close it. For the longer arc, the archive at /evolution-of-ai/ keeps this month in context.

Table of Contents