OpenAI released its new AI models GPT‑6.2 Astra and GPT‑6.2 Sol on Tuesday, both with new high benchmark scores and the former with a 20% price cut over the application programming interface (API) — so why are the new models drawing sharp criticism from several AI power users who tested them before launch?

Investor and former HyperWrite AI CEO Matt Shumer said he won't be releasing a longer written review as he has customarily done for other preceding OpenAI flagships. The new models "were essentially unusable,” Shumer wrote in a post on X. Developer and YouTuber Theo Browne also posted on X that 6.2 Astra was "the worst model OpenAI has ever released,” he wrote. They singled out the model's shortcomings in delegated work and creative tasks.

Wharton professor Ethan Mollick had a more mixed experience with Astra. It built a playable space elevator in Blender and Unity, but he found the model “very uneven & weird” and harder to steer than rival models.

Astra now costs $8 per million uncached input tokens and $40 per million output tokens while GPT-6.2 Sol keeps its predecessor 6.1 Sol’s $2 input and $10 output rates.

Both are proprietary, hosted models. They’re available now in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users, and through the OpenAI API as gpt-6.2-astra and gpt-6.2-sol.

OpenAI is positioning Astra for assignments such as changing a complex codebase, building a working web application, or producing a document that follows a company’s templates. Sol is the cheaper option for routine coding, writing, and workflows that require several steps.

The release comes eight days after rival Anthropic, in the run-up to going public, launched the well-reviewed Claude Fable 5.2 and Mythos 5.2, also at $8 and $40 per million standard input and output tokens, and a week after Google opened access to its own well-received Gemini 4 Carbon at introductory rates equal to Sol’s.

In other words, the competition is intensifying and it's probably one of the worst times for OpenAI to fumble, especially on the reception of its new flagship and middle-tier model.

What early testers said about writing, coding, and design

Dan Shipper, CEO of generative AI publication Every, said his team spent several weeks testing both models. He said Astra performed about as well as its predecessor on most of Every’s coding benchmarks, but often stopped early on end-to-end builds. He found that its push for brevity could make writing hard to follow or leave out necessary explanation.

He praised its default design instincts, but said its outputs were difficult to vary and had a “noticeable house style.” That distinction matters for teams that need one model to work across several brands or change the depth of an explanation for different readers.

In a follow-up post on X, Browne described how he and another tester independently asked Astra to design new levels for different games, Outer Wilds and Majora’s Mask. Both got water levels centered on music and bells, with the same name, “Bellwether.” The two runs show how similar Astra’s choices could be even when the projects differed.

Browne laid out even more objections in a long X post entitled “Why I hate GPT‑6.2 Astra,” saying it was the full text of a letter he had sent OpenAI during early testing.

On Browne’s experimental TypeScript compiler rewrite in Rust, he said Astra spent seven days without landing an improvement and produced a version that ran 56% slower after focusing on reducing its size. When he redirected it to improve speed, he said it spent another four days without landing a change.

Browne, did, however credit Astra with catching issues other models missed in code reviews and praised its 3D, game development, and computer-use capabilities. He said it helped him devise a reverse proxy that he expects will save him $25,000 a month. His criticism focused on how hard it was to keep Astra moving on long projects and understand what it had done afterward.

According to Browne, OpenAI addressed a few of his complaints before launch, and he still preferred Claude Opus 5.5 for his development work.

Shumer, for his part, said he gave Astra work to do as a sub-agent and the Claude models supervising it redid the work.

An X user posting as Jacob echoed Shumer’s dissatisfaction and speculated that it stemmed from how OpenAI trained the models to behave.

Mollick’s space elevator demo simulated weightlessness, according to Mollick, who said he still found himself reaching for other models.

In a LinkedIn comment on his post about the demo, he called this the first OpenAI frontier release since GPT‑4.5 that he did not see as a clear step forward and said he could still imagine 6.2 excelling at particular kinds of complex work. The launch-day posts reflect particular assignments and testers.

Other AI power users like Raindrop.ai CTO Ben Hylak and AI advisor Ashe Magalhaes shared more positive sentiments about GPT-6.2 Astra in particular, praising its strengths in animated visuals, motion graphics, VFX, and agentic AI video editing.

Meanwhile, OpenAI says many of its own employees already use Sol as their default for technical and other work.

How Astra and Sol compare on API pricing and access

At standard rates, Astra costs four times as much as Sol for the same number of input or output tokens. It matches Fable on both rates, though Anthropic charges half as much as OpenAI for cache reads on those flagships.

Model

Input ($/1M)

Output ($/1M)

Total ($/1M)

Source

Muse Spark 1.2 / 1.3 Contributor

$0.10

$0.20

$0.30

Meta

MiMo-V2.6-Flash

$0.14

$0.28

$0.42

Xiaomi

GPT-6 Luna

$0.10

$0.50

$0.60

OpenAI

Muse Ember — ≤200K input tokens

$0.10

$0.50

$0.60

Meta

Claude Haiku 5.5 — ≤100K prompt tokens

$0.10

$0.50

$0.60

Anthropic

DeepSeek-V4.1-Flash-1017 — off-peak

$0.15

$0.60

$0.75

DeepSeek

MiniMax-M3.1-Flash

$0.15

$0.60

$0.75

MiniMax

MiMo-V2.6-Pro

$0.435

$0.87

$1.305

Xiaomi

MiniMax-M3

$0.30

$1.20

$1.50

MiniMax

LongCat-2.0 — limited-time promo

$0.30

$1.20

$1.50

LongCat

DeepSeek-V4.1-Flash-1017 — peak hours

$0.30

$1.20

$1.50

DeepSeek

Muse Ember — >200K input tokens

$0.30

$1.50

$1.80

Meta

Claude Haiku 5.5 — >100K prompt tokens

$0.50

$2.50

$3.00

Anthropic

LongCat-2.0 — standard

$0.75

$2.95

$3.70

LongCat

Gemini 3.8 Flash — through Dec. 31, 2026

$0.75

$3.75

$4.50

Google

Muse Spark 1.1 / 1.2 / 1.3

$1.25

$4.25

$5.50

Meta

Muse Spark Open — Together AI serverless

$1.25

$4.25

$5.50

Together AI

GLM-5.3

$1.40

$4.40

$5.80

Z.AI

Grok 4.7 — ≤200K prompt tokens

$2.00

$6.00

$8.00

xAI

Qwen3.8-Max

$2.00

$6.00

$8.00

QwenCloud

Gemini 3.8 Flash — starting Jan. 1, 2027

$1.50

$7.50

$9.00

Google

Gemini 4 Hydrogen

$1.50

$7.50

$9.00

Google

GPT-6 Sol — Standard, ≤272K input

$2.00

$10.00

$12.00

OpenAI

GPT-6.2 Sol — Standard, ≤272K input

$2.00

$10.00

$12.00

OpenAI

Claude Sonnet 5.5

$2.00

$10.00

$12.00

Anthropic

Gemini 4 Argon — introductory pricing

$2.00

$10.00

$12.00

Google

Kimi K3.1 Pro

$2.00

$10.00

$12.00

Moonshot AI

Gemini 4 Carbon — announced introductory pricing

$2.00

$10.00

$12.00

Google

Grok 4.7 — >200K prompt tokens

$4.00

$12.00

$16.00

xAI

Grok 4.7 Fast (Cursor) — ≤256K input

$4.00

$12.00

$16.00

Cursor

GPT-5.4

$2.50

$15.00

$17.50

OpenAI

GPT-6 Sol — Standard, >272K input

$4.00

$15.00

$19.00

OpenAI

GPT-6.2 Sol — Standard, >272K input

$4.00

$15.00

$19.00

OpenAI

Claude Opus 5.5

$4.00

$20.00

$24.00

Anthropic

Grok 4.7 Fast (Cursor) — >256K input, up to 500K

$6.00

$18.00

$24.00

Cursor

Gemini 4 Argon — post-introductory pricing

$4.00

$20.00

$24.00

Google

Muse Wonder 1.0

$4.00

$20.00

$24.00

Meta

Gemini 4 Carbon — post-introductory pricing

$4.00

$20.00

$24.00

Google

GPT-6 Sol — Fast, ≤272K input

$4.00

$20.00

$24.00

OpenAI

GPT-6.2 Sol — Fast, ≤272K input

$4.00

$20.00

$24.00

OpenAI

Sakana Fugu Ultra (≤272K)

$5.00

$30.00

$35.00

Sakana AI

GPT-6 Sol — Fast, >272K input

$8.00

$30.00

$38.00

OpenAI

GPT-6.2 Sol — Fast, >272K input

$8.00

$30.00

$38.00

OpenAI

Claude Opus 5.5 — Fast mode

$8.00

$40.00

$48.00

Anthropic

Claude Fable 5.2 / Claude Mythos 5.2

$8.00

$40.00

$48.00

Anthropic

GPT-6.2 Astra — Standard, ≤272K input

$8.00

$40.00

$48.00

OpenAI

Claude Fable 5.1 / Claude Mythos 5.1

$10.00

$50.00

$60.00

Anthropic

GPT-6 Astra — Standard, ≤272K input

$10.00

$50.00

$60.00

OpenAI

GPT-6.2 Astra — Standard, >272K input

$16.00

$60.00

$76.00

OpenAI

GPT-6 Astra — Standard, >272K input

$20.00

$75.00

$95.00

OpenAI

GPT-6.2 Astra — Fast, ≤272K input

$16.00

$80.00

$96.00

OpenAI

GPT-6 Cyber

$20.00

$100.00

$120.00

OpenAI

GPT-6 Astra — Fast, ≤272K input

$20.00

$100.00

$120.00

OpenAI

GPT-6.2 Astra — Fast, >272K input

$32.00

$120.00

$152.00

OpenAI

GPT-6 Astra — Fast, >272K input

$40.00

$150.00

$190.00

OpenAI

GPT-6 and GPT-6.2 Sol cost exactly the same at each Standard/Fast and short/long-context tier, while GPT-6.2 Astra is 20% cheaper than GPT-6 Astra at the baseline — $48 versus $60 on this chart — with the savings carrying through the long-context and Fast tiers. OpenAI explicitly documents the same >272K multiplier and 2× Fast pricing structure for the older GPT-6 models.

Anthropic charges $10 per million tokens to write a five-minute Fable cache and $16 for an hour. OpenAI lists $10 for a 30-minute Astra cache and $20 for 12 hours at its short-context standard rate.

Google’s Carbon price sheet lists implicit caching only and includes thinking tokens in the output rate. Carbon’s December 1 price change moves it from Sol and Sonnet’s input and output rates to Opus’s.

The new Astra and Sol both have a 1.05 million-token context window and support up to 128,000 output tokens. For requests with more than 272,000 input tokens, OpenAI doubles input and cache rates and multiplies the output rate by 1.5 for the entire request. That puts a standard long-context Astra request at $16 per million uncached input tokens and $60 per million output tokens. The cutoff matters for agents that repeatedly send large repositories or document collections back to the model.

OpenAI’s model documentation lists half-price batch processing and charges twice the corresponding standard rate for Fast. Its faster Ultrafast API tier starts at three times standard pricing for both models and rises to six times on March 1, 2027, Pacific time. In ChatGPT Work and Codex, Pro 500 subscribers get Ultrafast on both models, while Pro 200 subscribers get it on Sol. Eligible Enterprise and Edu workspaces can enable access through an admin.

What the benchmarks say about the upgrade

In the nine head-to-head benchmark charts in OpenAI’s launch post, GPT‑6.2 Astra beats GPT‑6 Astra on all nine, and GPT‑6.2 Sol beats GPT‑6.1 Sol on eight of nine. The exception is Sol on the Vals Science Index.

These comparisons use each model’s best plotted score, so the reasoning setting can differ between runs. Independent Codex testing also found a flat result for Astra and a regression for Sol. OpenAI

Where the scores rose, stayed flat, or fell

Benchmark

Astra vs. GPT‑6 Astra

Sol vs. GPT‑6.1 Sol

Agents’ Last Exam — multistep professional work

Higher, 78.3% vs. 73.1%

Higher, 73.0% vs. 72.5%

GDPval v2 — professional deliverables

Higher, +6.2 points

Higher, +5.3 points, per OpenAI

Harvey LAB‑AA — legal work, strict pass rate

Higher, 15.0% vs. 8.6%

Higher, 11.1% vs. 6.9%

Terminal‑Bench‑Science — research workflows

Higher, 77.1% vs. 68.1%

Higher, 68.0% vs. 57.0%

Vals Science Index — scientific reasoning

Higher, 68.4% vs. 65.1%

Slightly lower, 58.5% vs. 59.6%

FrontierCode 2.0 — changes to real codebases

Higher, 44.2% vs. 42.8%

Higher, 43.8% vs. 39.0%

BenchCAD 2.1 Core — CAD tasks

Higher, 82.5% vs. 62.6%

Higher, 73.1% vs. 52.6%

AutomationBench — business workflows

Higher, 52.4% vs. 41.4%

Higher, 41.2% vs. 36.1%

LifeVal — everyday tasks

Higher, 86.1% vs. 79.8%

Higher, 83.8% vs. 79.5%

Artificial Analysis Coding Agent Index — Codex at max effort

Flat

Lower, −3 index points

The first nine rows come from OpenAI’s launch materials. The last is an independent test. Artificial Analysis traced much of Sol’s coding-agent decline to a 14-point drop on SWE‑Atlas‑QnA‑AA, where Sol more often left out key facts when answering questions about codebases. So both models improved on FrontierCode, while the broader Codex coding-agent result was flat for Astra and lower for Sol. OpenAI

OpenAI’s system card also shows mixed results. Both models improved on length-adjusted HealthBench Professional, while both scored lower on the raw version. OpenAI says most alignment results were similar to GPT‑6 Astra’s. Both new models took fewer improper actions in simulated workplace tasks, but more often failed to disclose obstacles that had blocked their work. OpenAI Deployment Safety Hub

GPT‑6.2 Astra is OpenAI’s highest-scoring model on the Artificial Analysis Intelligence Index, at 62, up five points from GPT‑6 Astra and second overall behind Claude Fable 5.2.

Epoch AI gives Astra and Fable the same point estimate of 174, compared with 169 for GPT‑6 Astra, with overlapping uncertainty ranges. OpenAI’s own composite also puts the new Astra ahead of its predecessor, 170.1 vs. 168.4. Individual tests can produce a different ranking. For example, OpenAI’s specialized Astra for Law configuration scores higher than base Astra on Harvey LAB‑AA.

On Terminal‑Bench‑Science 0.1, a 70-task test of research work in a terminal, OpenAI reports the new Astra gained nine percentage points over its predecessor at less than one-third the cost, reaching roughly 77% in its launch chart.

Artificial Analysis measured Fable at 73.3% at maximum reasoning with its mini-swe-agent setup. Anthropic reported 74.6% for Fable using Claude Code and 75.3% for Mythos. Those results put the models in reach of one another across the published runs, with different agent software and settings limiting what the gaps establish.

In legal work, OpenAI says Astra more than doubled Fable’s pass rate on Harvey LAB‑AA at roughly one-sixth the cost. The chart also puts Astra’s best result at about 15%. The benchmark only awards a pass when the work satisfies every grading criterion and avoids material hallucinations.

OpenAI’s separately configured Astra for Law adds a legal search index and research instructions and scored higher. ChatGPT Enterprise customers can request that configuration through OpenAI’s Trusted Access Program, including for API use.

Google and Anthropic compete at both price levels

Anthropic positions Opus 5.5 as its general-purpose model for coding and knowledge work, with Fable for harder assignments and Sonnet as a cheaper option.

That puts Sol at the same rate as Sonnet and half the price of Opus for base input and output. Fable’s optional fast mode doubles its input and output rates, giving teams a different speed and price trade-off from Astra’s Ultrafast tier.

Google’s lineup makes the same choice across a narrower price range. Google describes Carbon as its deepest reasoning model and says it uses more tokens than Gemini 4 Argon, even though the two share a rate card. Google points to strong Carbon results on CursorBench, WANDR, and AutomationBench, which cover software engineering, web research, and workflow automation.

Its Gemini 4 Hydrogen costs 25% less than Sol per standard input or output token and is aimed at faster execution. The Agents’ Last Exam leaderboard lists Hydrogen at 71.4%, using the Antigravity command-line agent, illustrating why workload and agent setup both matter when choosing among the cheaper models.

Security teams also face different access rules. OpenAI says both new models hit its Critical cybersecurity threshold and ship with stricter controls that can block some legitimate vulnerability research. At launch, its Daybreak program still points approved users to GPT‑6 Astra or GPT‑6 Cyber for authorized security work.

Anthropic uses the same underlying model for Fable and Mythos, with Mythos access available through its verification programs for eligible cyber and life sciences work. For teams evaluating these models for defensive security, the version they can put into production and the requests it will accept matter alongside the benchmark scores.

Open weights change the deployment decision

Moonshot’s Kimi K3.1 Pro gives buyers an option to run a model with downloadable weights on their own infrastructure. On Agents’ Last Exam, the published Kimi Code run scored 74.4% and passed 43.4% of tasks. The model has 2.8 trillion parameters in total, so the infrastructure needed to serve it is a material part of the comparison. Moonshot’s custom license permits internal use and fine-tuning.

It also imposes separate agreement and attribution requirements in specified commercial scenarios, including for large model-as-a-service businesses. Enterprises embedding Kimi in a product and businesses offering it directly through an inference API need to read those terms against their deployment plans.

DeepSeek V4.1 Flash 1017 offers weights under the MIT license and a much lower first-party hosted token price. Its published Agents’ Last Exam run scored 65.4% and passed 36.1% of tasks using OpenClaw. That still leaves a meaningful question for buyers with narrowly defined workloads.

If the model clears their acceptance tests for extraction, triage, or a bounded coding task, the price gap can pay for many more calls. Operating it in-house shifts the bill to accelerators, serving infrastructure, and the people running it. Using DeepSeek’s API keeps the per-token model and introduces a hosted provider into the data path.

The procurement comparison comes down to what it costs to get an output the business can use. For an agent that prepares a finance workbook, that includes the tokens, any data lookups, the runs that fail validation, and the time an analyst spends correcting the file.

A code agent adds test execution and review. Shipper’s experience suggests counting how often someone has to send a coding agent back to finish the original assignment.

Mollick’s demo suggests tracking how many rounds of direction it takes to get a usable result, even when that result is impressive.

OpenAI’s launch makes Astra cheaper at the meter and gives Sol a stronger case at its existing price. Whether either lowers the cost of finished work depends on how often it gets that work right in the software, with the permissions, and under the review rules the enterprise already uses.