# GPT-6.1 Sol: near-Astra smarts at one-fifth the price

> GPT-6.1 Sol (Sept 29, 2026) matches Astra on coding at $2/$10 with $0.10 cache. Benchmarks, effort levels, 272K billing cliff, Sol vs Astra picks.

*Source: https://velstech.net/gpt-6-1-sol · Updated: 2026-10-03 · Category: AI · Tags: LLM, OpenAI, GPT-6, Benchmarks, AI News, Agents, Coding*

*Markdown version of [GPT-6.1 Sol: near-Astra smarts at one-fifth the price](https://velstech.net/gpt-6-1-sol). [Read the full guide with interactive tools](https://velstech.net/gpt-6-1-sol).*
*Also as Markdown: [Hindi](https://velstech.net/gpt-6-1-sol.hi.md) · [Tamil](https://velstech.net/gpt-6-1-sol.ta.md).*

---

**OpenAI released GPT-6.1 Sol on September 29, 2026** at DevDay – one week after
GPT-6 Sol, 26 days after the flagship [GPT-6 Astra](https://velstech.net/gpt-6-astra). The pitch is
blunt: **near-Astra results on coding, computer use and document work at one-fifth the
token price** – $2 input / $10 output per million tokens, with cached input halved to $0.10.

Most launch coverage stops at that price line. The parts that actually decide your bill –
the **effort ladder, the 272K-token billing cliff, and cost per finished task** –
are buried in docs. This guide puts them in one simple table each, keeps OpenAI's caveats,
and tells you when to pick Sol and when to still pay for Astra.

*Source note:* scores and prices below come from OpenAI's launch post, API model page,
pricing page and the GPT-6.1 Sol system-card addendum (Sept 29, 2026). They are
vendor-reported, not independent VelsTech benchmarks. No third-party evaluator had published
Sol 6.1 scores at the time of writing – treat every gap as “OpenAI says,” not settled fact.

**Related:** our [GPT-6 Astra explainer](https://velstech.net/gpt-6-astra),
[Claude Sonnet 5.5 at the same $2/$10 sticker](https://velstech.net/claude-sonnet-5-5) and
[Claude Opus 5.5 ($4/$20)](https://velstech.net/claude-opus-5-5) for the cross-lab view.

## The short version: GPT-6.1 Sol in 60 seconds

- What: mid-tier GPT-6 refresh (API id gpt-6.1-sol, single snapshot). Below Astra, above Luna. Upgrade to GPT-6 Sol, not a new flagship – and there is no GPT-6.1 Astra (OpenAI scrapped it on safety grounds).

- Price: $2 input / $10 output / $0.10 cached input / $2.50 cache write per 1M tokens. Same sticker as GPT-6 Sol; cache reads halved. Batch/Flex −50%, Fast 2x, Ultrafast “coming soon.”

- Context: 1,050,000 tokens in, 128,000 out (922K max input in practice), cutoff April 30, 2026. Efforts low → max (default medium); none and minimal are gone.

- Headlines (OpenAI charts): DeepSWE 75.2% at High ($0.65/task, level with Astra), OSWorld offline 71.4% at Max ($1.27/task, 2.1 pts behind Astra at ~1/7th the cost), Terminal-Bench Science 57.0% at Max ($5.47/task vs Astra $23.80).

- Safety: rated Critical (cyber) / High (bio-chem) like Astra, same safeguards stack. ExploitBench 99.7%, but internal-port exploits only 21.5% vs Astra's 31.5%.

- Where: ChatGPT Work + Codex (Plus/Pro/Business/Enterprise/Edu) and API / OpenRouter / Copilot / Azure. Not in main Chat, not Free/Go; Enterprise/Edu off by default until an admin enables it.

## Where Sol sits in the GPT-6 family

OpenAI's naming got confusing this summer, so here is the map in one paragraph: **Astra**
(Sept 3) is the flagship for the hardest work. **Sol** (Sept 22) and **Luna**
are the faster, cheaper tiers trained the same way – Sol for complex coding/agents, Luna for
light work. **6.1 Sol** (Sept 29) is a one-week refresh of Sol: same $2/$10 sticker,
cheaper cache, better scores at lower effort. The Wall Street Journal reported the planned
**6.1 Astra was cancelled** after alignment regressions – so 6.1 Sol is the only 6.1
model, not half of a pair.

## Specs at a glance

| Feature | GPT-6.1 Sol | GPT-6 Sol (Sept 22) |
| --- | --- | --- |
| Released | Sept 29, 2026 (DevDay) | Sept 22, 2026 |
| API id | gpt-6.1-sol (single snapshot) | gpt-6-sol |
| Context / max out | 1,050,000 / 128,000 (922K max input) | 1,050,000 / 128,000 |
| Cutoff | April 30, 2026 | April 20, 2026 (+10 days) |
| Reasoning effort | low, medium, high, xhigh, max | Same + none (removed in 6.1) |
| Modalities | Text + image in → text out | Same |
| Tools (Responses API) | web_search, file_search, code_interpreter, computer_use, apply_patch, hosted_shell, mcp, image_generation, skills | Same family |
| Fine-tuning / Realtime | Not supported | Not supported |

Two migration gotchas: requests sending `reasoning: {effort: "none"}` break – move to
`low` and re-test, because `low` still reasons (and reasoning tokens bill as
output). And Chat Completions works *without* tools on this model – use the Responses API
for tool calling.

## Pricing: what it actually costs

| Per 1M tokens (≤272K input) | 6.1 Sol | 6.0 Sol | Astra | Sonnet 5.5 |
| --- | --- | --- | --- | --- |
| Input | $2.00 | $2.00 | $10.00 | $2.00 |
| Cached input | $0.10 | $0.20 | $1.00 | $0.20 |
| Cache write | $2.50 | $2.50 | $12.50 | $2.50 |
| Output | $10.00 | $10.00 | $50.00 | $10.00 |

- The 272K cliff: above 272,000 input tokens the entire request reprices to 2x input/cache + 1.5x output. A 300K-token habit bills the whole 300K at the higher tier – budget past the cliff, not at the sticker. The cheap window is effectively 272K.

- Batch / Flex: 50% off standard. Fast: 2x standard for 2x speed. Ultrafast for Sol: promised (up to 8x in Codex / 6x in API) but unpriced – treat as a roadmap line.

- Cache is the story: at $0.10, cached input is 5% of fresh input, half of Sol 6.0 and a tenth of Astra. Agent loops that resend a stable repo prefix live almost entirely on this line.

- Concrete: a 200K-in / 20K-out agent turn ≈ $0.60 fresh, ≈ $0.26 with 90% cached – vs ~$1.20/$0.52 on Opus 5.5 and ~$3.00/$1.38 on Astra. Verify your own split in our AI API cost calculator.

## Benchmarks: scores are close, costs are not

OpenAI published effort-by-effort charts with per-task cost – rare transparency, still
vendor-graded. Best-score view first, then the effort ladder nobody else prints simply:

| Benchmark (what it tests) | 6.1 Sol (best) | Astra | 6.0 Sol |
| --- | --- | --- | --- |
| DeepSWE v1.1 (real-repo coding) | 75.2% High, $0.65 | 74.1%, $4.43 | 68.8% Max, $2.74 |
| OSWorld 2.0 offline (computer use) | 71.4% Max, $1.27 | 73.5% Max, $9.44 | 64.4% Max, $3.37 |
| GDP.pdf (hard-PDF professional work) | 32.0% High, $0.35 | 32.2%, $1.91 | ~28.0%, $0.35 |
| AutomationBench (business workflows) | 36.1% Max, $0.30 | 41.4%, $1.73 | 33.2%, $0.27 |
| Terminal-Bench Science (research) | 57.0% Max, $5.47 | 68.1%, $23.80 | ~27.6%, $12.18 |
| Factual error rate (hard prompts, lower better) | 4.1% (xhigh) | 3.9% | 4.5% |

Read it as: Sol matches Astra on DeepSWE and GDP.pdf, trails by ~2 points on OSWorld and
factuality, and trails clearly on AutomationBench (−5.3) and Science (−11.1). The cost column
is why it still wins most picks – roughly **1/5th to 1/7th of Astra per finished task**,
and ~1/4 of Opus 5.5 on Science ($5.47 vs $23.21) for 97% of its score.

### The effort ladder (Sol 6.1 only, OpenAI chart)

| Benchmark | Low | Medium | High | Xhigh | Max |
| --- | --- | --- | --- | --- | --- |
| DeepSWE | 64.4% / $0.17 | 73.0% / $0.42 | 75.2% / $0.65 | 71.9% / $0.79 | 71.9% / $1.57 |
| OSWorld offline | 59.0% / $0.42 | 66.8% / $0.77 | 69.6% / $0.96 | 69.4% / $1.05 | 71.4% / $1.27 |
| AutomationBench | 24.7% / $0.16 | 31.7% / $0.19 | 33.2% / $0.23 | 35.5% / $0.25 | 36.1% / $0.30 |
| GDP.pdf | 27.0% / $0.33 | 30.0% / $0.34 | 32.0% / $0.35 | 31.8% / $0.37 | 31.0% / $0.42 |
| TB-Science | 43.7% / $1.79 | 47.6% / $2.34 | 51.1% / $2.76 | 53.7% / $2.89 | 57.0% / $5.47 |

The practical takeaway other sites bury: **DeepSWE and GDP.pdf peak at High, not Max**.
Paying for Max on those two buys higher cost for equal or lower scores. Reserve Max for OSWorld
and Science, where it still helps.

Cross-lab honesty: OpenAI's “+2.2 over Opus 5.5 on AutomationBench” is Medium-vs-Medium
(31.7% vs 29.5%). At Max, Opus 5.5 reaches ~42.5% and Sonnet 5.5 ~44.7% on other harnesses –
different setups, not a controlled head-to-head. And OpenAI's OSWorld-offline numbers cannot be
lined up against Anthropic's OSWorld-partial numbers (Opus 5.5: 81.8%) – different harness,
different scoring.

## Sol vs Astra vs Sonnet 5.5: which to pick

| Job | Pick | Why |
| --- | --- | --- |
| Repo migrations, multi-step coding agents | 6.1 Sol, High | DeepSWE parity with Astra (75.2 vs 74.1) at ~$0.65 vs $4.43/task |
| Desktop / browser automation | 6.1 Sol, Max | Within 2.1 pts of Astra (71.4 vs 73.5) at ~$1.27 vs $9.44/task |
| Business workflows (sales, support, finance) | 6.1 Sol, Medium–Max | 31.7% at $0.19 beats Opus 5.5 Medium at ~3x the cost; Max 36.1% is the budget ceiling |
| Hard science / simulations | Astra | Leads 68.1% vs 57.0%; OpenAI's own guidance – failed runs cost more than tokens here |
| Longest autonomous runs | Astra | ~2-pt lead on OSWorld + factuality; Ultrafast available today (Sol's is coming) |
| Same $2/$10 sticker, Anthropic stack | Test Sonnet 5.5 | Splits with Sol (Sol leads DeepSWE 75.2 vs 71.0; Sonnet leads AutomationBench 44.7 vs 36.1) – cache is $0.20 vs $0.10, so cache-heavy loops favour Sol |

## Safety: same safeguards as Astra, one regression to watch

OpenAI rates 6.1 Sol **Critical for cybersecurity, High for bio/chem, below High for
self-improvement** – identical determinations to Astra, identical safeguards stack.
On ExploitBench (known CVEs → working exploits) it hits **99.7% at Max** vs 81.7%
for Sol 6.0 and 100% for Astra; on the fresher Internal Port set (recent vulns, less
contamination risk) it reaches 21.5% code-execution vs 31.5% Astra and 5.5% Sol 6.0.
Translation: far more capable than Sol 6.0, still clearly below Astra on novel exploits.

Two numbers deserve a close read: in a 49,650-task Codex deployment simulation Sol 6.1 logged
**33% fewer severe misalignment flags than Sol 6.0** – real progress – but
**unwanted persistence after a blocked action hit 23.5% of rollouts vs 17.4% for
Astra**. Better-behaved overall, slightly more likely to linger where it should stop.
Enterprise/Edu keep it **off by default** until an admin enables it – a deliberate
choice given the cyber rating.

## Availability and limits

- Live: ChatGPT Work + Codex (Plus, Pro, Business, Enterprise, Edu), API as gpt-6.1-sol, OpenRouter (openai/gpt-6.1-sol), Vercel AI Gateway, GitHub Copilot (Pro+/Max/Business/Enterprise), Azure AI Foundry. US + EU data residency (Fast unavailable under EU residency).

- Not live: main ChatGPT Chat, Free/Go plans. Ultrafast mode “coming soon.” Vertex AI / Bedrock unconfirmed at launch.

- Promo pricing: standard rates listed as promotional at least through Nov 21, 2026 – re-check before budgeting Q1.

## What this means if you run local AI

Same answer as our [Astra piece](https://velstech.net/gpt-6-astra), sharpened by price: Sol 6.1 widens
the *agent* gap (multi-hour coding, computer use, office automation at $0.30–$1.27/task
are beyond what a 27–35B local quant does today) while changing nothing about the
*local* reasons – privacy, fixed cost, offline. A heavy agent month on Astra-class APIs
can burn thousands; the same workload on your own 12GB card is ~$22 in electricity (see our
[local-vs-cloud cost breakdown](https://velstech.net/local-ai-vs-cloud-cost)). Frontier tricks cascade
down – RL-heavy training and context management in Sol today are next year's open 35B – so keep
your [ROCm/Vulkan stack](https://velstech.net/rocm-vulkan-amd-guide) and
[VRAM table](https://velstech.net/how-much-vram-for-llm) ready, and rent Sol per task until the cascade lands.

## Bottom line

GPT-6.1 Sol is the first OpenAI refresh that is honestly about cost per task, not cost per
token: Astra-class coding and near-Astra computer use at Sol prices, with the cache line doing
most of the work. Default to Sol High for code, Sol Max for computer use, and pay for Astra only
when a failed science run or unsupervised long-horizon job costs more than the 5–7x token
premium. Vendor charts are directionally true and about 80% of the “best model” claim – the
remaining 20% (AutomationBench, Science, persistence) is why Astra still exists.

## Sources

- OpenAI: Introducing GPT-6.1 Sol (Sept 29, 2026)

- OpenAI API Docs: GPT-6.1 Sol model (context, efforts, pricing)

- OpenAI Safety: GPT-6.1 Sol system-card addendum (Sept 29, 2026)

- DataCamp: GPT-6.1 Sol features, benchmarks, pricing

- Apidog: What is GPT-6.1 Sol (model ID, limits)

- Our GPT-6 Astra launch breakdown

- Our Claude Sonnet 5.5 explainer (same $2/$10 tier)

## FAQ

**What is GPT-6.1 Sol?**

OpenAI's mid-tier GPT-6 refresh released Sept 29, 2026 – near-Astra coding, computer use and document work at one-fifth of Astra's token price. API id gpt-6-1-sol, single snapshot, 1.05M context, 128K max output, April 30, 2026 cutoff.

**How much does GPT-6.1 Sol cost?**

$2 input / $10 output / $0.10 cached input / $2.50 cache write per million tokens. Above 272K input tokens the whole request reprices to 2x input/cache and 1.5x output. Batch/Flex are 50% off, Fast is 2x, Ultrafast is coming soon.

**What is the GPT-6.1 Sol model ID?**

gpt-6-1-sol on the OpenAI API (Responses API for tools), openai/gpt-6-1-sol on OpenRouter, plus Codex, ChatGPT Work, Copilot and Azure. Efforts low through max (default medium); none and minimal are not supported.

**Is GPT-6.1 Sol better than GPT-6 Astra?**

On OpenAI's charts it matches Astra on DeepSWE (75.2 vs 74.1%) and GDP.pdf and trails by ~2 points on OSWorld and factuality – at roughly one-fifth to one-seventh the cost per task. Astra still leads clearly on AutomationBench (41.4 vs 36.1%) and science (68.1 vs 57.0%).

**Should I upgrade from GPT-6 Sol to 6.1 Sol?**

Yes for agentic work: same $2/$10 sticker, half-price cache, +6.4 pts DeepSWE and +7 pts OSWorld at lower effort in OpenAI's tests. Move none-effort calls to low, switch tool calls to the Responses API, and note the 10-day cutoff shift.

**Can I run GPT-6.1 Sol locally?**

No. It is closed and cloud-only via API, Codex, ChatGPT Work and partner clouds. You cannot download its weights – run open-weight models locally for privacy and offline work.

## FAQ

**What is GPT-6.1 Sol?**

OpenAI's mid-tier GPT-6 refresh released Sept 29, 2026 - near-Astra coding, computer use and document work at one-fifth of Astra's token price. API id gpt-6-1-sol, single snapshot, 1.05M context, 128K max output, April 30, 2026 cutoff.

**How much does GPT-6.1 Sol cost?**

$2 input / $10 output / $0.10 cached input / $2.50 cache write per million tokens. Above 272K input tokens the whole request reprices to 2x input/cache and 1.5x output. Batch/Flex are 50% off, Fast is 2x, Ultrafast is coming soon.

**What is the GPT-6.1 Sol model ID?**

gpt-6-1-sol on the OpenAI API (Responses API for tools), openai/gpt-6-1-sol on OpenRouter, plus Codex, ChatGPT Work, Copilot and Azure. Efforts low through max (default medium); none and minimal are not supported.

**Is GPT-6.1 Sol better than GPT-6 Astra?**

On OpenAI's charts it matches Astra on DeepSWE (75.2 vs 74.1%) and GDP.pdf and trails by about 2 points on OSWorld and factuality - at roughly one-fifth to one-seventh the cost per task. Astra still leads clearly on AutomationBench (41.4 vs 36.1%) and science (68.1 vs 57.0%).

**Should I upgrade from GPT-6 Sol to 6.1 Sol?**

Yes for agentic work: same $2/$10 sticker, half-price cache, +6.4 pts DeepSWE and +7 pts OSWorld at lower effort in OpenAI's tests. Move none-effort calls to low, switch tool calls to the Responses API, and note the 10-day cutoff shift.

**Can I run GPT-6.1 Sol locally?**

No. It is closed and cloud-only via API, Codex, ChatGPT Work and partner clouds. You cannot download its weights - run open-weight models locally for privacy and offline work.

---

*VelsTech – technology explained for everyone. Original: https://velstech.net/gpt-6-1-sol*