返回文章列表
aiNovember Releasesgpt62-work-package-routing
🧠

Two GPT-6.2 Models, One Engineering Work Queue

A release supervisor decides when GPT-6.2 Sol can prepare a repair and when Astra merits a second pass, using bounded engineering evidence, review minutes, and published API rates.

iBuidl Research2026-11-1816 min 阅读

TL;DR

OpenAI introduced GPT-6.2 Astra and Sol on November 17. Their API rates create room for a team to reserve Astra for work that merits a broader investigation. Route engineering requests by the evidence available at intake, the consequences of an incomplete change, and the time required for a person to verify it. A second model pass earns its place when it resolves a named uncertainty.

The operations team at a fictional distributor opens Monday with an awkward mix of software requests. A duplicated report heading is annoying six clerks. A migration to a new unit-of-measure field is holding up a supplier feed. A nightly job keeps missing its contracted cutoff, and three people have three different explanations. Each task can be described in a few lines. Their similarity ends when the team asks what would make a proposed repair acceptable.

This is the useful setting for the two releases. OpenAI positions GPT-6.2 Sol as the everyday model for work including coding and tools, and GPT-6.2 Astra as the flagship for more demanding projects. Its launch reports improvements across professional outputs and coding evaluations. Those results can inform which models a software team puts on its test bench. The results of the team's own acceptance tests decide whether a ticket is ready to ship. OpenAI's release announcement

08:35 — Three tickets arrive with different evidence

The supervisor reads the tickets with the team member who owns next week's release. She records the concrete defect, the affected workflow, and the smallest useful demonstration of a repair. To gather comparable evidence later, she preserves the first version of each description alongside the documents attached at intake. When an investigation turns up a missing rule, the team can see that it learned something during the work.

The report ticket comes with two archived examples, a screenshot, and a formatter test. The examples agree that the report should display one column title and still carry the same field in the downloaded data. The UI and download share a schema but generate their headings separately. The route into the code is narrow enough that a maintainer can enumerate the consumers before the work begins. A ticket like this is a suitable Sol candidate. The handoff asks for a patch and an explanation of which presentation output changed. The reviewer checks both outputs.

The supplier ticket contains a different kind of problem. The warehouse stores a count in individual units. One supplier sends an order in packs, and its new format carries the pack size in a second column. A different supplier's draft specification uses the same column for a case size. A message from the integration owner says the first feed is being migrated. The examples in the test folder bundle data from both sources. A broad automatic normalizer could produce plausible counts in every test while changing the quantity for a supplier outside the migration.

The supervisor holds that ticket at intake while the owner confirms the source boundary and the meaning of the two columns each supplier sends. The cheapest useful intervention is a question to a person with the contract. After that answer arrives, Astra can map which import, display, comparison, and export paths interpret the stored count. A second pass is valuable at that point because one unresolved risk spans multiple documented paths. The model gets a deliberately small input set with the relevant contract extracts and access to the approved repository snapshot.

The nightly-job ticket comes with timestamped spans showing that the schedule begins on time. Its owner suspects a recently added serialization step and can name the measurements that would separate CPU time from queued I/O. The first assignment asks Sol to build a reproducible timeline from those spans and to locate the relevant instrumentation. The team can take that packet to a deeper investigation if it cannot account for the gap at the end. An initial explanation of uncertainty is an acceptable deliverable on this ticket.

What the first reviewer receives

A useful handoff fits on one screen. It carries a stable revision identifier, an affected workflow, and a material claim about each observed change. A claim can cite an existing test, the freshly observed result of an allowed test, or an explicit assumption. A separate field records work still required from the release owner. The patch stays attached to the evidence packet so the reviewer can compare the exact revision with the recorded run.

The report owner compares two cells in the affected view and opens the download. The new screenshot can show the visual correction. A sample generated download can show that the field still belongs in the output data. The test command and its status let a maintainer repeat those checks later. A change to the schema would push the ticket back to intake because it would affect a wider set of callers than the assignment permits. The handoff makes that change visible even when the page looks good.

For the supplier migration, the reviewer starts with a five-order ledger. Two records are from the supplier in scope and preserve the link between the raw count, the pack size, and the canonical count. Two records are from the unchanged supplier and show the previous behavior. The remaining record has a missing size and should appear in an exception queue. In the fictional examples, twelve packs of eight become ninety-six individual units. A case-size column of eight on the unchanged feed has only the meaning already defined by that feed's own contract. The same number can support different decisions.

An Astra pass can review the ledger against boundaries in the assignment and look for a downstream consumer that assumes all quantities arrived in packs. It might return a shorter repair after identifying one shared function with too many callers. It might also return an evidence map that justifies splitting the work into a parser change and a compatibility check. The release job belongs to the maintainer who owns the import contract. That person accepts the behavior after inspecting it on the agreed sample.

The nightly-job investigation yields a different packet. It carries elapsed time for each stage and a cumulative total. If the reported stages consume twenty-six of the missing forty minutes, the reviewer has a fourteen-minute measurement problem. The investigation can record idle worker time, a delayed upstream acknowledgment, or overhead outside the traced segment as hypotheses. Replacing the suspected serialization stage before measuring the gap would consume review capacity and could leave the deadline failure intact.

How a second pass gets a purpose

An engineer writes one sentence authorizing the escalation. On the supplier ticket the sentence identifies whether exports regenerate quantities from raw partner fields after the import path has normalized them. This gives Astra a specific job and tells the reviewer what evidence ends the extra pass. The author can also give it the most credible alternative explanation so it can test both. Accepting a model's finding may require an ordinary instrumented run or a controlled replay with approved sample data.

A decline to escalate is also recorded. On the report ticket all requested behavior has been demonstrated and any residual uncertainty lies with a browser variant the maintainer can check directly. A second patch would add a review branch and another version of the same simple task. The team spends its time reproducing the view in that browser. The result then belongs to the same ticket regardless of which model first touched it.

The supervisor avoids using the model's reported confidence as the escalation switch. A model can be confident about an interpretation supported by an incomplete description. It can be cautious when every necessary artifact for a straightforward change is available. The trigger therefore comes from the work itself. A new consumer appears. Two independent records disagree. A side effect extends beyond the target workflow. A test exercises a simulated path while the missing behavior is observed under a real input contract. Each of those discoveries can justify additional reasoning or a person who can supply authoritative evidence.

One practical check is to give a second reviewer the release packet before showing the first review's conclusion. Ask the second reviewer to state what the tests actually establish. Agreement can support a decision when both explanations cite the appropriate evidence. If both passes depend on the same incomplete fixture, agreement adds little. The independent source of evidence may be a separate path inspection, a stronger test input, or a recorded requirement from the service owner.

11:20 — A price that fits inside the review budget

The published Standard short-context API prices are two dollars per million input tokens and ten per million output tokens for GPT-6.2 Sol. GPT-6.2 Astra is eight dollars for the same unit of input and forty dollars for output. The current pricing page offers additional terms for caching, processing tiers, and other features. The calculation here uses the simple uncached rate because the fictional team's weekly estimate starts with a small workload inside that pricing bracket. OpenAI API pricing

For example, assume one pass reports thirty-six thousand billed input tokens and eight thousand billed output tokens under the stated conditions. The arithmetic yields about fifteen cents on Sol and about sixty-one cents on Astra, before any other charges. The fourfold ratio comes from using identical quantities at rates that differ by a factor of four. Real requests can use different numbers of reasoning and other billed tokens, and documents or repeated tool calls can expand a prepared packet. The team needs measured usage for its own mix.

Now assume a queue of eighty assignments, each with that same toy usage. A Sol-first week would spend $12.16 on the first passes. If eight of those assignments also receive full Astra passes, the modeled amount rises to about $17.02. Giving every assignment a single Astra pass with that usage would cost $48.64. The simple difference is about $31.62. This example excludes review, failed attempts, tools, and incidental bills. It shows how the difference arises and which assumption to replace first after a trial.

The supervisor has another constraint. The team can spend forty reviewer minutes a day on model-assisted changes. A report patch takes three minutes to inspect. A migration investigation may need eighteen minutes because someone must follow two real consumption paths and consult the integration owner. At this scale the supervisor will usually feel a bad handoff in the review queue before the token bill. If universal escalation makes a reviewer read two long explanations for each ticket, the slower release cycle matters more than the illustrative savings.

The cost ledger therefore has three independent quantities. There is billed model work attached to all attempts. There is human review work attached to all proposed changes. There is outstanding work that someone must still resolve after the model concludes. A focused evidence packet can reduce the last two amounts even when it costs more to prepare. The trial compares total verified outcomes alongside the components of spending. It keeps the components visible so a team can negotiate the trade rather than argue about a total that disguises it.

Place deployment boundaries where engineering already enforces them

The team grants a work session the resources necessary for the ticket. It provides a known repository snapshot and the existing test harness for the report repair. The migration work uses sanitized data samples arranged by the integration maintainer. Access to a supplier's live service, publication rights, and permission to launch a production import belong to normal release controls. A reviewer can approve a revision through the same account and documented role used for other code changes.

This arrangement preserves a useful distinction in the launch documentation. Both API model pages describe support for text and images as input and text responses. They list tools that can be used with the Responses API. A model's ability to work with a file or a tool describes a supported interface. The application and its administrators decide what an invocation can access and where a write becomes effective. Teams should model their actual connected tools as part of the trial. GPT-6.2 Sol API model and GPT-6.2 Astra API model

One useful guard for ordinary maintenance is a separate release identity. A local work session can prepare a candidate and run appropriate checks. A maintainer can then apply the candidate through branch review and the test service. Its deployment identity inherits the existing approvals. The team can observe which stage attempted each operation without having to reconstruct authorization from a chat transcript days later.

Even a read-only investigation has scope. The work queue describes acceptable source material and the output destination. In the fictional supplier case, the integration owner can provide a redacted test packet that preserves units and the separate supplier contract. Old screenshots including account records or public issue attachments from unrelated work have no role in the sample. Good input selection improves both privacy and the reasonableness of the final result.

Remove easy cases from the model comparison

Early trial data can reward the choice of tickets instead of the model. If Sol handles the known presentation bugs and Astra handles uncertain migrations, a higher Sol pass rate says little about their relative capability on the same ticket. The supervisor can measure queue efficiency while keeping a separate matched set for comparison. The matched set gives each candidate identical facts, the same allowed actions, and the same acceptance conditions across a sample stratified by complexity.

A sample record includes the first acceptable patch, absent evidence, and review time. The assessor pays attention to false completion. A correct description of a missing supply contract is a better result than a convincing change justified by an invented convention. Similarly, a cautious human-facing question can be the correct output when the fixture has an ambiguous unit. A trial that awards points only for presenting finished code would penalize this useful behavior.

The maintainer adds an unfamiliar consumer to one migration case after the first trial to see whether the work package elicits an accurate dependency analysis on revised inputs. Another case removes an ambiguous setup step after a teammate points out the wording. Results for those changes belong to separately named runs. That way improvements from a better task packet are visible separately from changes following a model upgrade.

The evaluation uses equal stop conditions. For the deadline ticket a completed investigation accounts for the missing interval, or identifies the remaining gap and the best measurement for closing it. Each candidate gets the same wall-clock ceiling and tool envelope. A timed-out draft can still be scored for partial evidence, with the accepted-outcome rate preserving the distinction. A raw average of completion times should include a clearly disclosed denominator, especially when unfinished tickets pile up at the end of the run.

A route change in the middle of the shift

A team in production can receive an unexpected release dependency in the middle of an evaluation. The fictional distributor gets exactly that when an internal data consumer asks to move the supplier migration two days earlier. The owner can keep the agreed import semantics and expedite the review, or divide the work into changes that can be accepted separately. A deadline field in the routing request, even if accurate, provides no evidence that the earlier date is attainable.

The supervisor takes another look at the reviewer calendar. On Thursday the integration owner has ten minutes available, with twenty more on Friday. Moving the date advances the time at which approval is needed and reduces the available verification time. The team may still fit a walkthrough of the core import on Thursday by using an evidence packet that identifies the precise tested rows and the two downstream callers. If separate export verification requires the Friday window, a deployment that changes both behaviors will wait. This scheduling decision is supported by named people and available evidence.

The experiment also has to account for abandoned model work. Suppose a first attempt investigates supplier selection and a corrected assignment arrives after eight minutes. Treat the first attempt as a consumed resource and its output as a distinct record. The owner can retain a supported observation from it by attaching the exact artifact and describing which facts survived the scope correction. Discarding its cost and keeping its useful work would favor routes that require repeated prompting.

A different example comes from the deadline queue. The first pass made enough progress to identify an unused span label, then reached the time ceiling. The reviewer can ask for one observed trace from the relevant stage. She does not have to authorize the same whole investigation again. A better measurement reduces the next work packet and makes its result independently useful if the wider incident investigation takes another day.

When the initial choice proves expensive, the supervisor asks whether the missing resource was model capability, source evidence, review attention, or authority to act. Those categories inform a different remedy in each case. An accurate route can change mid-shift because the available evidence has changed. The ticket retains a short record of the reason and the state handed from one worker to the next.

The release lead's Friday decision

By Friday, the fictional team has enough information to change its intake. It defines a routine path for tickets with known affected behavior, firm boundaries, and three minutes of independent verification. It defines an investigation path for requests with conflicting requirements or multiple consumers. In the second path, the next useful step can be resolving the contract, asking Astra a prepared question, adding instrumentation, or assigning a colleague with access to evidence.

The team also assigns an exit condition to every investigation. The warehouse migration exits when the expected quantity is supported at import and at both named outputs on the same records. The overnight-job analysis exits when the final missing interval has an owner and a measured explanation. The report ticket exits after the corrected screen and intact export are observed. Those conditions make it possible for the supervisor to tell whether a smaller trace and cleaner evidence packet beat a larger one.

A useful early metric is time from intake to a reviewable packet. It reveals how much of the week's delay occurs before a model sees a complete request. A separate metric records time from reviewable packet to accepted work. It reveals when a candidate causes rework, particularly if a plausible explanation obscures a contract gap. The supervisor can compare these numbers over several weeks while tagging new classes of work or shifts in incident load.

Only then does she add the model names to the routing policy. The firm constraints in the rule are an intact test contract and a named reviewer. The model preference reflects current performance and can change after the next trial. The possibility of a better model is already accounted for in a process that can replace its candidates without changing what completed engineering work means.

Sources

更多文章