AI Brief, 18 November 2026: GPT-6.2's five scores, and a 220B nobody downloaded
OpenAI published GPT-6.2 Astra and GPT-6.2 Sol at 17:45 UTC on 17 November; the Hacker News
submission landed a minute later and finished the
day on 1,164 points and 733 comments, the loudest item of the window by an order of magnitude.
Astra's price falls to $8 per million input tokens and $40 per million output, 20% below GPT-6
Astra, with cached input down from $1.00 to $0.40. Sol holds at $2 and $10. Both carry a
1,050,000-token context window and reject any request setting reasoning effort to none.
The hardest fact in the release is that GPT-6.2 Astra does not have a score. Artificial Analysis, which measures rather than announces, published five on the day, one per reasoning-effort setting, from 45.65 at Low to 62.14 at Max — a 16.5-point spread at a thirteenfold difference in cost. ARC Prize published ten configurations, in which the same model solves 70.31% of ARC-AGI-3 under ARC's Standard harness and 100.00% under its Provider Adapter harness, rank 32 of 143 against rank 1 of 143, same benchmark, same day. The widely shared "100%" is the second figure. Astra's top-effort score puts it second to Claude Fable 5.2, 62.14 to 64.56, and on the Coding Agent Index it is flat while Sol loses three points.
The release's most consequential document is not the launch post. OpenAI's system card records that both models meet the Critical threshold for cybersecurity under its Preparedness Framework, and that Astra meets the High threshold for AI self-improvement — in OpenAI's words the first model it has deployed externally to do so. It also discloses regressions on benign-biology overrefusal, mental-health handling and self-harm responses.
Two independent papers published in the window also deflate the weight-space programming result this account led on 3 November: an outside evaluation of the released checkpoint finds it level with its unedited base. And two substantial open-weight releases landed almost unnoticed — Tencent's 220-billion-parameter Hy4-flash has thirteen likes.
- GPT-6.2 Astra and Sol, 17 November, 17:45 UTC. Astra $8/$40, 20% below GPT-6 Astra; cached input $0.40; 1,050,000-token context.
- One model, five scores: 45.65 to 62.14. The last 0.58 index points cost 45 times as much per point as the first 11.35.
- ARC-AGI-3: 70.31% or 100.00%, depending on harness — and the cheaper harness scores higher, $4,112 against $24,169 for the run.
- Astra is the first externally deployed model OpenAI places at High for AI self-improvement; both are at Critical for cybersecurity.
- Independently Astra is second to Claude Fable 5.2; on coding it is flat and Sol regresses three points. Both hallucinate at 59% and 55% at max effort.
- The MIT weight-space edit scores 73.9 against its base model's 73.5 on ARC in an outside evaluation, and below it on SQuAD.
- Tencent Hy4-flash: 220B total, 12B active, 1M context, open weights, thirteen likes and zero downloads.
The configuration is the model
This is an API release, with no parameter count or architecture detail published anywhere, and the cheaper model has the fresher cutoff: 1 June 2026 for Sol against 18 May 2026 for Astra. One pricing feature appears in no coverage: above 272,000 input tokens the input and cache rates double, so a million-token window is billed at one price across a quarter of its range and another across the rest.
The figures below are Artificial Analysis's own measurements.
Read that as a price list and two things follow. The marginal cost of an index point rises 45-fold across the dial: Low to Medium buys 11.35 points for 12.6 cents, about 1.1 cents a point, while Xhigh to Max buys 0.58 points for 29.2 cents, about 50 cents a point. And some settings are dominated by cheaper ones: Sol at Max scores 57.26 for 23.5 cents against Astra at Medium's 57.00 for 18.9 cents, equal capability for 24% less, from the dearer model. Running the cheap model hard is worse value than running the dear one gently.
The ARC numbers make the point more starkly, because the variable there is scaffolding rather than effort. ARC's Standard harness lets a model carry forward notes it chooses to keep; its Provider Adapter harness preserves opaque reasoning state between requests, so the model can reuse prior work. On ARC-AGI-3, Astra scores 70.31% at max effort under Standard, for $24,168.91, and 100.00% at xhigh effort under Provider Adapter, for $4,112.40 — thirty points and a sixfold saving, from changing how state is carried between turns. It is the second consecutive Astra release to bury that gap: ARC's 3 September post had GPT-6 Astra at 62.7% Standard against 99.9% Provider Adapter. A model that "solves" ARC-AGI-3 without a named harness is not a result.
OpenAI concedes the mechanism: asked on Hacker News why its comparison did not cite Artificial Analysis's coding scores, a staff member said that version "uses the mini-swe-agent harness, which can understate performance of models trained to use particular harnesses". One caveat belongs with all of it — on Artificial Analysis's hallucination measure, Astra and Sol guess wrong 59% and 55% of the time at max effort when out of their depth.
For a 20,000-token reusable prompt that is $0.008 a request cached against $0.16 uncached — at ten thousand requests a day, $80 against $1,600. Nothing errors. The fix is structural:
# GPT-6.2 appends top-level `instructions` AFTER input, so a reusable
# block passed this way now falls outside the cacheable prefix.
client.responses.create(
model="gpt-6.2-astra",
instructions=REUSABLE_PROMPT, # $8/M, every call
input=[{"role": "user", "content": question}],
)
# Put the reusable block first, inside `input`, and it is a prefix again.
client.responses.create(
model="gpt-6.2-astra",
input=[
{"role": "developer", "content": REUSABLE_PROMPT}, # $0.40/M cached
{"role": "user", "content": question},
],
instructions=note_for_this_call_only,
)
GPT-6.2 also discards reasoning items after mid-conversation edits to the prefix, with OpenAI warning that answer quality can fall and the model may redo work.
The weight-space edit, measured by strangers
On 3 November this brief led on Sparks of Weight-Space Programming, in which GPT-6 Astra altered 131,072 of Qwen3.5-0.8B's weights with no gradients and reportedly took its MATH-500 score from 7.9% to 48.1%, for 2,184 Hacker News points. The deflation here was that the baseline was broken, and a placebo control has been asked for since. Both arrived on 17 November.
Hyunjong Ok and Jaeho Lee at POSTECH published Weight-Space Programming Doesn't Need Transformers, whose main argument is elsewhere. The relevant part is an appendix: they downloaded the checkpoint the MIT group released and evaluated it — ARC 73.9 against the unedited base's 73.5, MMLU 47.0 against 46.7, MathBench 49.6 against 49.2, SQuAD 82.3 against 83.1 — three gains under half a point and one loss, across four test sets of a thousand items or more. In 87 of their own editing runs, every checkpoint already scoring 72.3% or above came out unchanged, and editing damaged one encoder from 63.7% to 33.1%. They do not address MATH-500 or the token cap. What they report is that the published artifact does approximately nothing.
The control comes from a second paper, submitted at 05:18 UTC on 17 November by a group across Fudan, the Shanghai Institute for Mathematics and Interdisciplinary Sciences, Tencent and Northwest University, on weight-space editing of Stable Diffusion rather than language models. It is preregistered, protocol dated 9 November, and includes the missing arm: budget-matched random search. On their primary measure the unit means are random search 75.0, the model-written program 74.3, a hand-written baseline 74.7; against random search the program scores −0.7, interval −9.0 to +8.0. Their summary is that the method shows no advantage over the strong controls. The domain is image editing, so this is not the MIT experiment repeated. It is the design that experiment lacked, and the arbitrary edit wins.
Two open-weight releases nobody noticed
Tencent published Hy4-flash at
2026-11-17T03:44:46Z with FP8 and base variants, 46 minutes before yesterday's edition, which did
not carry it. Nearly 25 hours on, it has thirteen likes and zero downloads, no English-language press and no
Hacker News thread.
It is 220 billion parameters with about 12 billion active per token across 42 layers, at a hidden size of only 3,072, with a 1,048,576-token window. Hybrid attention — 12 sparse layers against 30 carrying a 512-token window — plus experts routed in a latent space take the KV cache to one twenty-fourth of Hy3's and one sixth of Hy4's, which makes that window affordable rather than nominal.
The unusual choice is that 63.5 billion of the 220 billion parameters are an n-gram embedding table — learned bigram and trigram representations fetched by eight hash heads per order into a 10.3-million-entry vocabulary per head, at layers 1, 11, 21 and 31. Nearly 29% of the model is a lookup table the router never sees, which readers of the 16 November edition will recognise: Qwen4-Max carries 328 billion of what Alibaba calls N-gram Embedding parameters. Two Chinese labs, two days apart, both counted a very large n-gram table in a headline parameter figure — one more reason that figure no longer compares across labs.
Every figure is Tencent's own: on WorkIndex, its internal productivity suite, Hy4-flash scores 42.1 against 51.4 for Hy4, at a stated twentieth of its cost per task. The licence is Tencent's own, not an open-source one.
SenseNova-Loop-9B, from SenseTime, arrived
at 2026-11-17T10:16:41Z under Apache 2.0, with 215 likes and a Hacker News submission at two
points and no comments. The repository is gated
behind a name-and-email form, so the figures here come from the project's GitHub
mirror of the same card.
It is a weight-tied depth-recurrent transformer, a class long present in papers and almost never in a shipped checkpoint. Of its 32 decoder layers, layers 0 to 3 run once,
layers 4 to 27 run twice with identical weights, and 28 to 31 run once. Writing that block as
so depth for computation is 56 while depth for memory is 32: 9.41 billion stored parameters doing what SenseTime calls 14.6 billion parameters' worth of work.
The loop count is one config field, loop_settings.eval_count, fixed rather than adaptive — which
leads to the release's most interesting admission. SenseTime's own loop-scaling ablation finds the
best results at three or four loops, not two: against a single pass, three loops gain up to
16.93% on one task set and four up to 8.55% on another, and going from the shipped two to
four adds a further 1.04 and 1.35 points. Two ships "to balance inference efficiency", a
throughput decision published beside the evidence that it is not the accuracy-maximising one — so
anyone holding these weights can buy accuracy with a one-line change and twice the compute. Its figures are self-reported against the three-times-larger Qwen3.8-27B, and it loses NL2Repo 41.36
to 44.60. There is no technical report and no paper, engine support needs SGLang #47281 or vLLM #64832 by hand, and the promised official AWQ build does not exist — in its absence the
community produced roughly twenty conversions inside 24 hours, around a gated repository.
Also notable
- Microsoft launched MAI Code 2 and MAI Code 2 Flash at 18:15 UTC on 17 November, comparing itself against GPT-6.1 Sol rather than the model that shipped thirty minutes earlier. API-only, no weights, no published token price. Its model card names the harness: MSBench via GitHub Copilot CLI 1.0.102, subagents disabled, three runs averaged. But read the chart data behind the launch page and MAI Code 2 leads only one of its four benchmarks, FrontierCode v1.1, the one it did not measure itself, losing SWE Atlas to Claude Opus 5.5 by 14.4 points. Its "40% of the cost" claim also uses a 50% promotional price.
- NVIDIA published a quantised Qwen3.8-Flash-Next on 17 November, 129.8 GB, after 100 optimizer updates of quantisation-aware distillation from a frozen BF16 teacher. Five of its six benchmark rows move by under a point, and the card states the gains "are not statistically established" — a vendor shipping a quantisation and declining to claim it helped.
What to watch
- Whether anyone runs SenseNova-Loop-9B at four loops and publishes it. The vendor's own ablation says it should be better, the knob is one config field, the cost is twice the compute, and the vendor has already said which way it expects the answer to go.
- Whether ARC-AGI-3 results start carrying a harness column. Two consecutive Astra releases have produced a thirty-point spread between two scaffolds, and the cheaper one wins. Until the harness travels with the number, one ARC score is not comparable to another.
- Whether the MIT group responds to its checkpoint being evaluated at parity. The POSTECH appendix is checkable by anyone, since the diff, the checksums and the base model are public. The honest resolutions are a MATH-500 rerun under a raised token cap, or a withdrawal.
- Whether anyone ablates an n-gram embedding table. Two flagship models now spend 63.5B and 328B parameters on one, and the experiment is cheap for whoever holds the weights: drop the Engram lookups at layers 1, 11, 21 and 31 and report the perplexity delta. Hy4-flash's weights are downloadable.
- Still open: Mistral has published nothing for 34 days, day 29 of the Voxtral 3 watch; no independent score for Inkling 2 Small, flagged on the 13th, where the 16th brought throughput rather than a benchmark; no independent token count behind the Qwen4-Max efficiency claim, asked on the 16th; and no threshold-conditioned fail-open rate from any decision-model vendor, day 40, with Liquid AI's d1.2 adding a fourth endpoint on the 17th.