Evidence note: claims from the supplied brief

The benchmark results, demonstrations, safety ratings, prices and November 17 launch details below come from the brief. Sublyx has not independently verified them or tested either model. Check your account’s catalog for current access and billing before making a request.

What changed against the baselines?

Start with the comparison, not the leaderboard position. On FrontierCode, for example, the claim is about getting close to a reference score for less money. It is not a claim that Astra beat the reference.

BenchmarkReported result versus baselineReported cost versus baseline
Agents Last ExamGPT-6.2 Astra: +5.2 points over GPT-6 Astra1/3 of GPT-6 Astra's cost
GDPvalGPT-6.2 Astra: +6.2 points over GPT-6 Astra; GPT-6.2 Sol: +5.3 over GPT-6.1 SolNo run cost quoted
FrontierCodeWithin 0.5 points of an unnamed reference1/5 of reference cost
LifeValAbove GPT-6 Astra; margin not supplied1/7 of GPT-6 Astra's cost

LifeVal's one-seventh figure describes the cost of that evaluation, not the price of a token. The quoted token rates show a smaller reduction. That distinction matters when estimating the cost of a job both models can finish in one call.

Astra: judge the finished artifact

Documents and analyst work

The brief reports an AA AnalystAgent score of 62.5% across all five attempts. Keep the attempt condition beside the score; it is not a first-response accuracy figure. For a document trial, give Astra two revisions of the same budget and ask it to reconcile the totals. Check each changed line against its source page. A polished explanation that mixes last quarter's figures with this quarter's is still a failed reconciliation.

On Harvey LAB-AA v1.1, the brief credits Astra with more than twice Claude Fable 5.2's pass rate at roughly one-sixth the cost. The Basis comparison instead reports timing figures of 50% and 73%. Those are measures of time, not legal accuracy: for a document review, time the work through sign-off, including the minutes a reviewer spends checking citations.

CAD, 3D and robotics

The brief puts GPT-6.2 Astra at 1318 Elo on AA-3D, ahead of Claude Fable 5.2 at 1271 and GPT-6.2 Sol at 1259. These are relative ratings, not percentages of usable CAD files. The examples include a pendulum clock, a 1:8 nitro RC buggy and a chassis simulation. Each calls for a different check: does the clock mechanism move without interference, do the buggy parts fit, and do the simulation’s load assumptions match the design?

For robotics, distinguish describing a movement from completing it within constraints. Test Astra's proposed actions in simulation against the same starting states as Sol's. Count collisions and failures to stop separately from tasks left unfinished. Do not give a model direct actuator access on the strength of a benchmark.

Code, reasoning and science

FrontierCode suggests a useful coding trial: compare accepted patches under a fixed budget, not lines of code generated. Keep build failures and retries in the bill. The reported 100% on ARC AGI3 used an adapter; the result describes that combination, not an unaided chat session.

On Terminal Bench Science, the brief reports about 78% at under $10. It does not specify enough billing detail to turn that into a per-experiment quote. To test a scientific workflow, ask for a plot and the script that generated it from a fixed dataset. Rerun the script in a clean environment and check units and missing values before evaluating the written interpretation.

The brief also names the Vals Science Index and a $300 Erdős problem. A benchmark result and a prize problem are different kinds of evidence: for the latter, the useful artifact to review is the proof, not the dollar amount.

Cybersecurity and safety

The brief describes Astra's cyber capability as critical, mentions Daybreak, and rates self-improvement capability as high. It also claims a 35-second kill switch. That timing is not a guaranteed shutdown deadline for an integration. For a defensive review, give the model read-only access to a sanitized configuration and ask it to identify excessive permissions. Keep approval and revocation outside the model’s control.

Sol: speed, context and suitable tasks

The brief puts Sol's Intelligence score at 57.3 versus 55. In its robotics comparison, Astra scores 79, Sol 78, Claude Fable 5.2 75 and Gemini 4 Hydrogen 58. If Sol meets the accuracy requirement on your task, the more useful distinction may be the time a user spends waiting.

The brief reports an observed 2–3× speedup for Sol, cites 112 tokens per second for its predecessor and gives a task-time comparison of 8 minutes versus 19 and 26.2 minutes. Do not multiply 112 by the speedup to estimate Sol's throughput: elapsed task time includes reasoning and tool waits as well as token generation. With dynamic reasoning and caching, test both a fresh request and one with a repeated prefix, and time each through the first usable answer. The reported codename “Bell” or “Bel” is unconfirmed.

The stated 1.05M-token context breaks down as 922k input + 128k output. Do not budget 1.05M tokens of documents and then add a response on top. Instructions, conversation history and tool results also consume input space. Confirm the limits on the exact route you intend to call.

Try Sol first for ticket classification, extracting fields from a manual, or answering a narrow question about a long release history. For the last task, place an obsolete policy near the beginning and its replacement near the end. Require the answer to identify the effective date and quote the current passage. Fitting the text is useful only if the model retrieves the right version.

Pricing: compare the invoice, not the score

API prices quoted in the brief, USD per million tokens
ModelInputCached inputOutput
GPT-6.2 Astra$8$0.40$40
GPT-6 Astra$10$1.00$50
GPT-6.2 Sol$2$0.10$10
GPT-6.1 Sol$2$0.10$10
Claude Opus 5.5$15Not stated$75

At those rates, 100,000 uncached input tokens and 10,000 output tokens would cost $1.20 on GPT-6.2 Astra, $1.50 on GPT-6 Astra, $0.30 on either Sol and $2.25 on Claude Opus 5.5, before tools or retries. Sol's quoted rates are unchanged; Astra's ordinary input and output rates fall 20%. The brief also mentions Batch and Flex. Check their rates and turnaround terms separately if a job can wait.

Launch plans and model selection

The brief dates the launch to , across ChatGPT Work, Codex and the API. It names Plus, Pro, Enterprise and Edu plans and distinguishes Work from regular Chat, which it says was not part of the rollout. It also names Amazon Bedrock and places Sol's ultrafast mode on Pro 200 and Pro 500. A plan listing does not establish that a particular account, region or API route has access.

If both become available, start with Sol for interactive, easily checked work and Astra for multistep tasks where rework is costly. Run both on a held-out sample with the same acceptance checks. Promote Sol only if it meets your quality floor within the response deadline; escalate ambiguous cases to Astra only if the second pass improves enough results to justify its cost. Keep the previous route available until the pilot meets both targets.

Try one catalog-confirmed model

The brief suggests gpt-6.2-astra and gpt-6.2-sol as candidate IDs, not guaranteed Sublyx identifiers. On a trusted machine, set SUBLYX_API_KEY in the environment and set SUBLYX_MODEL_ID to the exact ID copied from the Sublyx model catalog. Confirm chat completions support for your account. Run this Python 3 example once per candidate; it uses the standard library and sends no files or tool requests.

import json
import os
from time import monotonic
from urllib.request import Request, urlopen

payload = {
    "model": os.environ["SUBLYX_MODEL_ID"],
    "messages": [{"role": "user", "content":
        "March budget: $80. Revised March budget: $95. "
        "Return only the change in dollars."}],
}
request = Request("https://api.sublyx.org/v1/chat/completions",
    data=json.dumps(payload).encode(), headers={
        "Authorization": "Bearer " + os.environ["SUBLYX_API_KEY"],
        "Content-Type": "application/json",
    })
start = monotonic()
with urlopen(request, timeout=60) as response:
    result = json.load(response)
print(result["choices"][0]["message"]["content"])
print("Seconds:", round(monotonic() - start, 2), "Usage:", result.get("usage"))

The expected change is an increase of $15. A correct answer checks the route and one small task, not the headline benchmarks. If the request fails, check authorization and the catalog entry before retrying. Never paste the key into a support transcript.