Some of the hardest activity successful codification reappraisal happens extracurricular the changed lines. A alteration tin look correct successful isolation and still break codification elsewhere successful the system.
That is what makes our early results for OpenAI's GPT-6 Astra astir interesting. In our evaluation, Astra caught astir 4% much branded bugs done actionable findings than GPT-5.6 Sol, and 22% much than Opus 5.
The biggest jump comes connected harder cross-file reviews, wherever Astra's gains scope 20% complete Sol and 33% complete Opus 5. Using that capacity astatine customer standard besides intends protecting customer information and assessing the model’s nationalist API pricing.
What Astra added to codification review
Our measurement present is actionable bug coverage, meaning really galore branded bugs a exemplary catches done findings a developer tin enactment on.
Coverage is rounded to 1 decimal; comparative gains usage unrounded values. The wide summation complete GPT-5.6 Sol appears humble successful this first information measure. This information besides includes simpler reviews, wherever location whitethorn beryllium little room for a stronger exemplary to differentiate itself; Astra's larger advantage appears successful the harder cross-file subset. It is an early, directional result.
Coverage is rounded to 1 decimal; comparative gains usage unrounded values. Compare models wrong this chart, since its reappraisal trouble differs from the wide evaluation. The harder cross-file comparison is much encouraging. Astra's comparative advantage grows to 20% complete Sol and 33% complete Opus 5. That suggests worth successful connecting a change's intent to consequences distributed crossed a codebase.
These results picture 1 portion of reappraisal performance. They do not found an wide ranking of reappraisal quality, foretell a team's defect rate, aliases committedness the aforesaid summation connected each propulsion request.
Putting discourse to work
A ample discourse model creates room for information. Useful reasoning requires the exemplary to place which pieces matter, link them, and scope a conclusion supported by evidence.
Our mentation is that Astra's astir absorbing beforehand lies successful connecting the correct information. Its larger gains connected harder cross-file reviews propose advancement connected activity wherever applicable accusation is distributed. They do not isolate the origin of that advancement aliases beryllium that much discourse unsocial improves a model's performance.
OpenAI positions Astra for multistep activity crossed code, browsers, and master software.
For teams building pinch models, the useful mobility is wherever the other reasoning changes the result capable to warrant its cost. A difficult task pinch scattered grounds is simply a stronger campaigner to analyse than a regular task that a little costly exemplary already handles reliably. That doesn't mean switching each convention to Astra, maxing retired reasoning effort, and letting it rip.
Stronger reasoning comes astatine a premium
Astra’s modular API rates are $10 per cardinal input tokens and $50 per cardinal output tokens, according to Astra’s published pricing. Fable 5.1 has the aforesaid guidelines input and output rates, though caching prices differ. Anthropic’s pricing documentation provides the afloat breakdown.
To put those prices successful context, see an schematic task utilizing 100,000 uncached input tokens and 10,000 billable output tokens, including reasoning tokens. Holding token usage changeless makes the published rates easier to compare; existent task costs alteration pinch usage.
All figures usage publically listed modular API prices, checked September 4, 2026. OpenAI figures usage short-context rates. Sol's nationalist promotional pricing is disposable astatine slightest done November 21, 2026. The illustration excludes caching, cache writes, tools, retries, location uplifts, and service-tier adjustments. Real tasks tin usage different numbers of tokens. At that fixed usage, Astra costs 2.5 times Sol, astir 4.7 times Terra, and astir 47 times Luna. Those are meaningful premiums. They are not predictions of the quality successful costs per completed task. A exemplary that needs less tokens aliases less attempts could constrictive the gap.
OpenAI reports little estimated task costs for Astra successful immoderate of its ain evaluations contempt higher token prices. That makes full costs per successful result worthy measuring connected your work, alternatively than assuming that either token value aliases a capacity people settles the decision. Read OpenAI's ratio guidance.
What mightiness transportation beyond codification review
The transferable thought is reasoning complete relationships among abstracted sources. Our findings propose respective uses worthy evaluating; we person not measured Astra connected these tasks:
- Research synthesis: reconcile conflicting reports, link claims to their evidence, and place gaps that a summary of each archive would miss.
- Operational investigation: combine a coherent mentation from logs, incident notes, and runbooks, while separating observations from hypotheses.
- Requirements and argumentation analysis: trace a projected alteration crossed specifications, soul policies, and implementation plans to emblem inconsistencies for master review.
- Document and spreadsheet work: cheque whether assumptions, formulas, and communicative conclusions work together crossed a study and its supporting materials.
The communal building is scattered grounds pinch limitations betwixt its parts. Start pinch bounded activity whose reply tin beryllium checked. The easiest measurement to trial whether Astra is correct for your workflows aliases products is to tally it alongside your existent exemplary connected the aforesaid tasks, past comparison reply quality, verification time, and full cost.
Building and balancing NIGHTSHIFT
We besides utilized Astra to build a full game: NIGHTSHIFT, an action RPG made pinch Godot and GDScript. Its hardest problem was balancing the interactions betwixt systems, past revisiting that equilibrium arsenic the crippled changed.
NIGHTSHIFT, built pinch Astra done quality guidance and iteration. That meant reasoning crossed 7 characteristic classes, a 988-node passive accomplishment tree inspired by the legendary Path of Exile passive node tree, progressive skills, runes and socketable upgrades, accomplishment evolutions, and co-op. Across 40 zones successful 10 acts, the run introduces progressively analyzable force swarms and combinations. Changing 1 people tin besides alteration which upgrades are useful, really a accomplishment develops, and what a statement tin handle.
We made basal changes to halfway systems during improvement and asked Astra to activity done the consequences and rebalance them. "Balance" successful a crippled for illustration this is the astir difficult imaginative situation for compelling gameplay. How tin you make sweeping changes aliases present wholly caller crippled systems and mechanics while keeping progression, combat, and trouble coherent?
We besides wanted players to observe overpowered endgame builds done clever combinations of classes, stats, items, passives, and progressive skills. The situation was shaping a progression wherever reaching moreover mid-game was uncertain, but productivity and experimentation could salary disconnected successful spectacular ways. The reward for uncovering those combinations was moving up to a build that could satisfyingly melt force swarms.
The process meant returning to the crippled betwixt different work, giving Astra feedback, and giving it afloat autonomy to usage its judgement successful refining that balance. Astra besides built autochthonal PS5 and Xbox controller support, autochthonal macOS, web, and Linux builds, and co-op. Co-op was peculiarly interesting, because Astra was capable to build it retired for playing connected the aforesaid compute and LAN, which since we wanted to administer to coworkers moving macOS, it required generating a caller Xcode project, App Store Connect account, found aggregate certificates and entitlements, and notarization.
Astra handled it each autonomously, only pausing to occasionally inquire for authority and permissions it didn't already have. As a nosy bonus, the crippled was besides built for agents themselves to beryllium players, which was somewhat surreal to beryllium capable to play unrecorded co-op sessions pinch Astra arsenic a teammate. Several of america recovered it difficult to put down. "I'm moving connected exemplary evals" became a useful mentation for having a crippled unfastened erstwhile a head stopped by.
The crippled gave america a imaginative mounting to research the aforesaid capacity that stood retired successful codification review. Astra had to logic astir really each alteration affected the remainder of the system.
What we want to spot next
The adjacent breathtaking beforehand would make this extent of reasoning dependable capable to usage much often. That intends accordant gains connected difficult work, conclusions group tin verify, and little full costs for a successful outcome.
That calls for amended action of applicable context, clearer supporting evidence, and less unnecessary steps. It besides intends privacy-preserving deployment paths that make precocious capabilities usable nether existent customer commitments.
Astra gives america an encouraging awesome connected cross-file reasoning. The applicable opportunity is to move that capacity into activity that is some much useful and easier to trust. As pinch bringing immoderate caller exemplary online astatine CodeRabbit, the information is 1 portion of that decision.
Protecting customer data
Neither CodeRabbit nor our exemplary providers train AI models connected CodeRabbit customers' proprietary codification aliases individual accusation collected during backstage codification reviews. Read our privacy policy.
OpenAI and Anthropic person different data-retention policies:
- OpenAI: GPT-6 Astra supports zero information retention for eligible API customers. OpenAI’s API information controls picture ZDR eligibility and supported capabilities. Astra announcement, OpenAI information controls.
- Anthropic: Fable requires 30-day retention by default for information monitoring, but eligible customers tin now usage Fable 5 and 5.1 pinch ZDR while Enterprise Frontier Safeguards is being introduced. For products serving different businesses, that action requires position agreed pinch Anthropic. EFS is designed to support retained activity information successful customer-controlled infrastructure. Covered-model retention policy, Enterprise Frontier Safeguards, eligibility details.
Any exemplary we usage for customer reviews must meet our data-protection requirements.
English (US) · 
Indonesian (ID) · 
