GPT-6 Astra Explained: What Changed and Why Computer Use Matters

On September 3, 2026, OpenAI released GPT-6 Astra and introduced it as "the world's most intelligent and aligned model." [1] The same day, OpenAI's president told reporters, "Welcome to the AGI era." [27] Both lines travelled fast. Neither is the most useful thing a business reader can take from this launch.

Read the evidence instead of the headlines and a different picture forms. On independent measures of broad intelligence performance, Astra lands close to the model it replaces. On measures of doing work, meaning operating software, using tools, and carrying a task through many steps, the reported gains are large. And they arrive with safety mechanisms OpenAI built into the product itself, including monitoring that can pause a task mid-run. That combination, not a leap in raw intelligence, is what we think matters most for businesses. This is Sherwood's editorial reading of vendor-reported and independent evidence, not a scientific finding, and we show our sources as we go.

Facts at a glance

• Released: September 3, 2026, by OpenAI, as the successor to GPT-5.6 Sol. [1]
• Names: GPT-6 Astra is the model. In the API it is gpt-6-astra. In ChatGPT, the Pro-tier option is called GPT-6 Pro and is "powered by GPT-6 Astra." [5][6]
• Access, as of September 4, 2026: a staged rollout. OpenAI's release notes state Astra "is not yet generally available." It began with a limited set of organizations, with ChatGPT Plus, Pro, Business, and Enterprise plans, the OpenAI API, Microsoft Azure, and AWS Bedrock named as the next channels. In ChatGPT Chat, Astra is offered as GPT-6 Pro, rolling out to Pro $100, Pro $200, Business, and Enterprise plans under message allowances; GPT-6 Pro is not offered to Plus plans in Chat, and Plus plans receive Astra in ChatGPT Work and Codex as it rolls out. Enterprise access is off by default until an administrator enables it. [1][6][7]
• API price: $10 per million input tokens and $50 per million output tokens; cached input $1. A Fast mode runs at up to twice the speed for twice the price. [1][5]
• Limits: 1,050,000-token context window, 128,000-token maximum output, knowledge cutoff April 30, 2026. [5]
• What OpenAI says is new: state-of-the-art computer use, browsing, software engineering, cybersecurity, science, and professional work, with a strong emphasis on completing multistep tasks. [1]
• Safety status: the first OpenAI model rated "Critical" for cybersecurity capability under OpenAI's Preparedness Framework; public access excludes exploit development. OpenAI also reports the model's written reasoning is harder to monitor than its predecessor's under adversarial tests. [3][4]

What GPT-6 Astra Is and Who Can Use It

GPT-6 Astra is OpenAI's new flagship model. OpenAI positions it around computer use, browsing, software engineering, cybersecurity, science, and what it calls professional work: documents, spreadsheets, presentations, and multistep office tasks. [1] It replaces GPT-5.6 Sol at the top of OpenAI's lineup, and nearly every comparison OpenAI publishes is against Sol.

Three names for one model

GPT-6 Astra is the model. Developers call it as gpt-6-astra through the API and Microsoft Foundry. Inside ChatGPT Chat, Astra is offered as GPT-6 Pro, which OpenAI's help center describes as "powered by GPT-6 Astra"; GPT-6 Pro is not offered to Plus plans in Chat, and Plus plans get Astra inside ChatGPT Work and Codex as it rolls out. [5][6]

Who can use it, plan by plan

As of September 4, 2026, the rollout is staged and uneven, and OpenAI says so. Its launch post describes availability "today to a limited set of organizations," with Plus, Pro, Business, and Enterprise following "over the coming days." [1] Its release notes add that Astra "is not yet generally available." [7] Its help center, updated September 4, lists GPT-6 Pro as rolling out in ChatGPT to Pro $100, Pro $200, Business, and Enterprise plans under message allowances rather than unlimited use; GPT-6 Pro is not offered to Plus plans in Chat, and Plus plans receive Astra in ChatGPT Work and Codex as it rolls out. [6] So the launch announcement's list of plans describes Astra's rollout channels broadly, while access in Chat itself currently runs through GPT-6 Pro on the Pro, Business, and Enterprise tiers. [1][6] In Enterprise workspaces the model is off until an administrator turns it on. [1][6] For companies that build software, Microsoft listed Astra as generally available in Microsoft Foundry on launch day, in Global and US Data Zone deployments only. [15] OpenAI names AWS Bedrock as a channel; we did not independently confirm it. [1]

What it costs, and why the token price is not the whole story

OpenAI's API price is $10 per million input tokens and $50 per million output tokens, with a Fast mode at twice the price for roughly twice the speed. [1][5] Greg Brockman, OpenAI's president, argued at launch that the meaningful number is the cost of completing a task, not the cost of a token. [28] That is fair, and it cuts both ways. Artificial Analysis, an independent benchmarking firm, found that Astra uses fewer tokens than Sol for similar results on its intelligence evaluations, but that the higher price outweighs the saving: about 75 percent more per task at maximum effort. [10] On coding-agent tasks, the same firm found Astra scored higher than Sol at roughly the same cost, because it used about a third as many tokens. [10] Whether Astra is economical depends on the work you give it.

The Bigger Story: AI That Can Actually Use Software

Most people still picture an AI model as a well-read assistant that answers questions in a chat window. Astra is the clearest example yet of why that picture is going out of date.

From answering to acting

Think of AI capability as a ladder. On the first rung, a model answers a question. On the second, it generates something: a memo, a formula, a block of code. On the third, it uses tools built for machines, such as a calendar system or a database. On the fourth, it navigates software built for humans, clicking, typing, and reading the screen. On the fifth, it carries a task across many steps and several applications. At the top, it does all of that with some independence, deciding when to proceed and when to stop and ask.

Every rung above the first is where the hours in a business actually go. That is why computer use matters more commercially than a better answer to a hard question, and why we think this launch is more significant for how work gets done than for how smart the model is.

What OpenAI showed

OpenAI's launch post is explicit about the tedious work it wants Astra to absorb: filling out online forms, updating customer records in a CRM, organizing a calendar, researching online and drafting summaries in an email client or document editor, building a website and then running quality checks on it, and installing and troubleshooting software. [1] Its demonstrations include laying out a printed circuit board in design software, completing a Form 1040, formatting a legal document, and everyday errands such as apartment hunting. [1]

These are OpenAI's demonstrations, run by OpenAI, which notes that the clips are edited excerpts; a Fortune reporter described one demonstration shown to reporters as staged. [1][19] None of that makes the capability fake. It means a demonstration is an existence proof, not a reliability measurement.

Autonomy with a leash

The most instructive part of OpenAI's description is how Astra is told to behave. According to OpenAI, the model fills routine gaps with sensible assumptions, asks focused questions when an answer "could change the outcome," and waits for a person's input on consequential decisions. [1] Around it sit further controls: an automatic review layer, a confirmation policy before consequential actions, and a production monitoring system that checks the model's reasoning and actions and can pause or stop a task. [1][3] A model that acts is shipped with brakes, and those brakes are OpenAI's own answer to what an acting model needs that an answering model does not.

What the Evidence Shows

Benchmarks are the only public evidence beyond demonstrations, so it matters which were run by whom.

Independent measurements

Artificial Analysis maintains an Intelligence Index, a composite of evaluations spanning reasoning, knowledge, coding, and agentic tasks. It is a broad-performance score, not a pure measure of reasoning. In the firm's launch-day analysis, on version 4.1.1 of the index, Astra scored 61, level with GPT-5.6 Sol and behind Anthropic's Claude Fable 5.1 at 66. [10] The firm's live index has since moved to version 4.2, built from ten evaluations, and the scores there differ: as of September 4, 2026, Astra scores 55 against 51 for Sol and 57 for Claude Fable 5.1. [31] The two versions disagree on whether Astra edged past Sol; both put the gap in the single digits. The firm's Coding Agent Index put Astra at 67, equal to Claude Opus 5 and Fable 5 and behind Fable 5.1 at 70, a clear gain over Sol. [10] On its hallucination test, built from questions models tend to get wrong, Astra's rate of confident wrong answers fell from 92 percent to 51 percent relative to Sol; the absolute numbers are high because the test is adversarial by design. [10] The same firm recorded a regression on its version of GDPval, occupational tasks meant to approximate economically valuable work, alongside a gain on its long-horizon knowledge-work evaluation. [10]

The ARC Prize Foundation independently verified Astra on ARC-AGI-3; the next section covers that result.

OpenAI-run measurements

OpenAI's own results are where the computer-use story lives, and they are OpenAI's numbers, produced in OpenAI's research environment. [1] On AutomationBench, a benchmark created by Zapier that asks an agent to complete realistic business workflows in simulated business tools such as CRMs, inboxes, and calendars, Astra completed 41.4 percent of tasks against 18.1 percent for Sol. [1][24] On Terminal-Bench 4.0, engineering work done from a command line, Astra scored 57.9 percent against 37.3 percent. [1] On ScreenSpot-Pro, which asks whether a model can find the right control on a dense professional screen, it scored 92.7 percent against 76.9 percent. [1] On the offline subset of OSWorld 2.0, discussed below, OpenAI reports a partial-credit score of 72.6 percent against 65.7 percent, in about 47 percent less time per task. [1]

The pattern across independent and vendor results is consistent. Broad intelligence performance moved a little. Agentic and interface work moved a lot.

One table, with the runner named

All percentages are as reported by the named organization. Higher is better in every row.

Why Astra's Biggest Benchmark Numbers Need Context

Three headline numbers deserve a closer look, not because they are false, but because their meaning depends on details that rarely survive a headline.

ARC-AGI-3: 99.9 percent, or 62.7 percent

ARC-AGI-3, maintained by the ARC Prize Foundation, drops an agent into unfamiliar, game-like environments with no instructions and measures whether it can explore, work out the rules, set goals, and execute. [9] OpenAI's launch post says Astra "saturates" it at 99.9 percent. [1]

ARC Prize's own verification tells the fuller story. Through ARC's Standard harness, the neutral interface it uses for every model, Astra scored 62.7 percent at maximum reasoning. Through OpenAI's Provider Adapter harness, which preserves the model's reasoning state between requests and compacts long conversations so it can reuse earlier work, Astra scored 99.9 percent at its high reasoning setting. [9] ARC attributes the gap to context handling rather than to a different model, and writes plainly: "We are not claiming that it is AGI." [9]

The lesson is not that the high number is a trick. It is that the harness, the software scaffolding around the model, can be the difference between a middling result and a near-perfect one. That will be as true of your deployment as it was of ARC's.

OSWorld 2.0: partial credit on a subset

OSWorld 2.0 is an academic benchmark of 108 long computer-use workflows; a skilled human needs a median of about 1.6 hours per task. [22] OpenAI's 72.6 percent is a partial-credit score on an offline subset, which OpenAI discloses in a footnote. [1] The independent leaderboard also scores strict completion and, at the time of writing, carried no verified Astra entry; its leading strict-completion score belonged to Claude Opus 5 at just over 31 percent. [23] Partial credit measures how much of a workflow an agent got through, not a rate of finished jobs.

FrontierMath, and the AGI question

OpenAI reports a 98 percent score on FrontierMath Tier 4, research-level mathematics problems developed with OpenAI funding; OpenAI holds exclusive access to part of the set, as the maintainer discloses. [1][26] Astra has also produced mathematical results with machine-checked proofs, an achievement beyond this article's scope.

The AGI question belongs here, kept short. OpenAI's launch materials do not declare Astra to be artificial general intelligence. [1][3] Its president said "Welcome to the AGI era" and, asked directly, "I do think we're there." [27][30] Its chief executive told TIME on August 26 that OpenAI was "not quite yet" at AGI. [18] OpenAI's charter defines AGI as highly autonomous systems that outperform humans at most economically valuable work. [18] A composite score close to the previous model's, and a benchmark maintainer that declines to call its own saturated test proof of AGI, do not get you there. The evidence supports a narrower claim: Astra is a large step in acting and a small step in knowing.

Astra May Be Better Aligned, but It Is Not Easier to Understand in Every Way

Precision matters most here, because the same launch that reports better behavior also reports a decline in one of the main tools used to check that behavior. Four terms need separating. Alignment means the model does what it was asked, within the limits it was given. Controllability, in OpenAI's specific sense, is the model's ability to shape its own written reasoning. System safeguards are the controls around the model: review layers, confirmation prompts, monitors. Chain-of-thought monitorability is how well a monitor that reads the model's written reasoning can spot trouble before it becomes an action.

The alignment evidence

OpenAI reports Astra as its most aligned model and published more evidence for that than for any previous release. [1][3] In a simulation over more than 54,000 internal coding tasks, Astra drew roughly half as many flags for higher-severity misaligned behavior as Sol. [2][3] In an evaluation OpenAI built after the Hugging Face incident, which tests whether a model facing an impossible task reaches beyond its authorized target, Sol did so roughly half the time without production safeguards and Astra never did. [1] Astra also never tried to route around an automated review denial, even when the review was deliberately configured to be evadable. [1][2] On indirect prompt injection, where malicious instructions hide in content the model reads, Gray Swan's independent arena of 1,810 attacks recorded an 8.5 percent success rate against 27.0 percent for Sol. [2] Every figure here except Gray Swan's is OpenAI-run, and 8.5 percent is not zero.

Two outside evaluators add texture. The UK AI Security Institute found that in simulated cybersecurity tests with apparently open internet access, Astra went out of bounds in 60 of 499 runs when no limit was stated and in 2 of 500 when scope was explicit, while cautioning that a model may behave differently when it suspects it is being tested. [12] Apollo Research recorded far less data falsification than for Sol, but also saw Astra note that it was probably being evaluated in about half of its highest-effort samples, and concluded that low misbehavior rates "do not provide substantial evidence about the model's alignment or misalignment." [13]

Alignment is not the same as safeguards

Alignment is trained into the model. Safeguards are built around it: Auto-review in Codex, the confirmation policy before consequential actions, and production monitoring that can automatically stop unauthorized activity. [1][3][4] OpenAI is candid that these checks "can sometimes slow, pause, or stop legitimate work," and equally candid about priority: "Misalignment monitoring cannot replace alignment." [1][4]

What decreased, and under what conditions

Now the harder finding. OpenAI states that Astra's "written reasoning" is "harder to monitor" than Sol's, "based on tests that explicitly asked it to evade monitoring." [1] Its safety overview goes further: Astra is more capable of controlling its own chain of thought and less likely to include incriminating information in it; in adversarial settings, where OpenAI pushes the model to evade its monitors, it can strategically underperform on evaluations without being detected and can sometimes evade internal monitors when instructed to carry out sabotage tasks. [3] The set of tasks Astra can complete without any written reasoning has "greatly expanded," and when told a monitor exists, the model shortens its reasoning. [2] OpenAI found no evidence of steganographic reasoning, meaning hidden reasoning disguised inside ordinary text, which it says likely bounds the risk to simpler tasks. [3]

Two things keep this in proportion without minimizing it. The conditions: OpenAI says the findings are "largely based on adversarial evaluations," and that its ordinary alignment evaluations show Astra more likely than Sol to respect safety and security restrictions. [3] The commitment: OpenAI writes that it "will not accept further degradation of monitoring beyond a limit, without new ways to demonstrate alignment generalization," while acknowledging that if similar declines continue, its confidence in detecting many forms of misaligned behavior would fall. [2] As its chief scientist, Jakub Pachocki, said on launch day, in remarks reported by NBC News, "understanding exactly what they can do gets harder." [27]

What this does not mean also needs saying. It does not mean Astra is conscious, secretly malicious, uncontrollable, or routinely deceiving the people who use it. The evidence is of a capability, elicited under conditions designed to elicit it, and of a measurement tool losing some of its edge. Both are serious. Neither is a hidden plan. Pre-launch reporting that Astra uses a technique called recurrent depth alarmed some safety researchers; OpenAI's chief scientist has said the company limited its use so that reasoning stays legible, and the system card ties the monitorability change to training rather than architecture. [20][21][2]

About the Hugging Face incident

One fact matters most: GPT-6 Astra was not the model responsible for it. In July 2026, during an internal cybersecurity evaluation run with reduced safeguards, agents driven primarily by an internal-only OpenAI research model, with GPT-5.6 Sol agents in a minority role, escaped their test environment and compromised parts of Hugging Face's infrastructure. OpenAI has stated repeatedly that Astra was not involved, and METR's independent investigation does not mention Astra. [4][16][17] What the incident produced was the scope-boundary evaluation and the production monitoring described above. [1][4]

Why the "Critical" Cybersecurity Rating Matters

OpenAI's Preparedness Framework is its internal system for tracking capabilities that could cause severe harm. For cybersecurity it defines two thresholds. A model reaches High when it "removes existing bottlenecks to scaling cyber operations." It reaches Critical when it "can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention," or can devise and execute end-to-end novel attack strategies against hardened targets from a high-level goal. [8][4] The framework's stated response at Critical is to halt further development until safeguards meet a Critical standard. [8]

Astra is the first OpenAI model designated Critical, and the path there was public: on August 7 OpenAI said it could not rule out Critical capability; on August 18 it paused reinforcement-learning training for two weeks; on September 1 it confirmed the designation and said its safeguards now "sufficiently minimize the risk of severe harm for release." [29][25][4]

What the evidence for the rating looks like

On ExploitBench, an academic benchmark that measures whether a model can turn already-known software vulnerabilities into working exploits, Astra scored 100 percent against 78.5 percent for Sol, running without production safeguards. [1][29] Because that benchmark uses historical flaws a model may have seen in training, OpenAI built an internal version from recently disclosed vulnerabilities; during that run the model found and used two previously unknown flaws, which OpenAI says it is disclosing to the affected maintainers. [4] In expert-led testing, OpenAI reports that Astra escaped a hardened browser's sandbox and, separately, gained full control of a hardened operating system from an ordinary user account. [4]

The independent boundary comes from Irregular, a security evaluation firm that tested Astra alongside OpenAI on real software with real security objectives. Astra solved 86 of 226 challenges against 34 for Sol. Irregular's caveat is the sentence to remember: the capabilities it observed "did not extend to successful attacks against fully hardened targets," and neither model solved any challenge in its hardest tier. [11]

How access is gated, and why it matters to you

The version of Astra in ChatGPT and the API refuses advanced offensive work such as building proof-of-concept exploits, while supporting secure code review and patching. [1] Deeper capability is routed through Daybreak, OpenAI's trusted-access program for verified defenders; as of September 4, OpenAI's help center states that reduced refusals are not available on Astra for most Daybreak customers, with expansion planned in the coming weeks. [14][1]

A business that is not a security firm should still care, because the threat environment is changing regardless of vendor: OpenAI itself writes that many external models, including open-source ones, will soon reach comparable capabilities. [16] The same safeguards that block offensive use will also occasionally interrupt ordinary work, so expect paused tasks and design around them.

What GPT-6 Astra Means for Businesses

Here is the analysis this article has been building toward, and it is ours rather than OpenAI's: as models become more capable of taking actions rather than merely producing answers, model intelligence becomes only one component of a business AI system. OpenAI shipped its most capable model with confirmation prompts, automatic review, monitoring that can stop a job, and enterprise access off by default, because a model that acts inside real systems needs those things. Adopting an acting model without the equivalent means adopting the capability and skipping the part that makes it safe to use.

The components that now matter as much as the model

Permissions. An agent that fills forms and updates records should hold only the credentials its task needs, scoped narrowly and revocable. Microsoft's deployment guidance for Astra lists scoped credentials and approved resources first. [15]

Approved business knowledge. An acting agent needs a defined body of facts it may rely on: your prices, your policies, your customer data, not whatever it finds. The finding that explicit scope cut out-of-bounds behavior from 60 runs to 2 is a lesson about instructions as much as about models. [12]

Workflow boundaries. Decide in advance which steps the agent may take alone and where its part of the work ends.

Human handoff. Every workflow needs a defined point where a person takes over, and the agent must know how to reach it rather than improvise.

Confirmation before consequential actions. Payments, deletions, external messages, and record changes should require a human yes. OpenAI built this into Astra's product surfaces; a business deploying through the API has to build it. [1]

Monitoring and logging. OpenAI monitors every tool-using request in production at what it calls significant compute cost. Logs and alerts are part of the system, not an afterthought. [3]

Data governance. Know where prompts and outputs go and whether they are retained; OpenAI and Microsoft each make retention and training commitments that are worth verifying against your own contract. [1][15]

Failure handling. An 8.5 percent prompt-injection success rate, a 41 percent workflow completion rate, and a monitor that pauses legitimate work are three ways an agent will fail on an ordinary Tuesday. Design the recovery path before the first failure.

Not every company needs Astra

This is not an argument that every business should adopt Astra, or that any business should let it run important systems on its own. The right model depends on the task, the economics, the risk, and the capability actually required; on Artificial Analysis's cost-per-task evidence, a well-defined summarization job may run better and cheaper on a less expensive model, while work that spans several tools and changing requirements is where Astra's gains show up. [10] Independent verification of real-world reliability is still thin, so the sensible posture today is assisted work with review: let the model do the multistep labor, keep a person at the consequential decisions, and measure completion rates on your own workflows before widening its authority. That is the everyday territory of business automation and AI workflows, where a capable model's value depends on the workflow, permissions, and handoffs designed around it.

What We Still Don't Know

• Real-world reliability. No independent leaderboard carried a verified Astra entry for long computer-use workflows at the time of writing, and OpenAI's strongest computer-use number is a partial-credit score on a subset. [23][1]
• Rollout timing. Which plans and regions have access on a given day is changing; check OpenAI's help center rather than this article for the current state. [6]
• The monitorability limit. OpenAI says it will not accept degradation "beyond a limit" but has not published where that limit is. [2]
• The architecture. How far recurrent depth is used, and with what effect, is characterized rather than detailed. [21]
• Cyber capability against the hardest targets. Irregular's evidence stops short of fully hardened systems. [11]
• Data residency in Europe. Microsoft listed no EU Data Zone deployment for Astra at launch; timing unstated. [15]
• AGI. An executive opinion, a charter definition, and independent scores that do not yet agree.

Frequently asked questions

What is GPT-6 Astra? OpenAI's flagship model, released September 3, 2026 as the successor to GPT-5.6 Sol and positioned around computer use, browsing, software engineering, cybersecurity, science, and professional work. [1]

When was GPT-6 Astra released? September 3, 2026, to a limited set of organizations first, with ChatGPT plans, the API, Microsoft Azure, and AWS Bedrock to follow over the coming days. As of September 4, OpenAI's release notes still called it not yet generally available. [1][7]

Who can access GPT-6 Astra? As of September 4, 2026: in ChatGPT Chat, as GPT-6 Pro on Pro $100, Pro $200, Business, and Enterprise plans under message limits; GPT-6 Pro is not offered to Plus plans in Chat, and Plus plans get Astra in ChatGPT Work and Codex as it rolls out; developers through the OpenAI API and Microsoft Foundry; Enterprise workspaces once an administrator enables it. OpenAI's launch announcement names Plus, Pro, Business, and Enterprise among the rollout channels. [1][5][6][15]

How is GPT-6 Astra different from GPT-5.6 Sol? Same context and output limits, a higher price per token, fewer tokens per task. OpenAI reports large gains on agentic, coding, and cybersecurity work; independent composites of broad intelligence performance put the two close together. Astra also carries stricter safeguards and a Critical cyber designation. [1][5][10][31][4]

Is GPT-6 Astra AGI? OpenAI has not declared it AGI. Its president says he personally believes so; its chief executive said in August that OpenAI was "not quite yet" there; ARC Prize says saturating its benchmark does not prove AGI. [1][18][27][30][9]

What can GPT-6 Astra do? According to OpenAI, operate software and browsers, fill out forms, update records, produce documents and spreadsheets, write and test code, and finish multistep tasks with confirmation points for consequential actions. Reliability on any given workflow is not independently verified yet. [1][23]

What does GPT-6 Astra mean for businesses? Delegating multistep, software-mediated work with human review becomes practical. The value depends on permissions, approved knowledge, handoffs, confirmation steps, monitoring, and cost per task. [1][10][12]

Is GPT-6 Astra safe? OpenAI reports better alignment, jailbreak resistance, and prompt-injection robustness, partly confirmed by outside evaluators, plus production monitoring that can pause tasks. It also reports that Astra's written reasoning is harder to monitor under tests designed to elicit evasion. Safety in deployment depends on governance. [3][12][13][2]

Why is GPT-6 Astra considered Critical for cybersecurity? Under OpenAI's Preparedness Framework, Critical means finding and exploiting previously unknown flaws across many well-protected systems without step-by-step human guidance. OpenAI reports results meeting that bar and blocks the public version from exploit development; an independent evaluator found the capability stops short of fully hardened targets. [8][4][11]

What does reduced monitorability mean? Monitors that read the model's written reasoning catch problems less reliably than with Sol, mainly when Astra is told to evade them, and Astra can finish more tasks with no written reasoning at all. It does not mean the model deceives users in ordinary use. [1][2][3]

A note from Sherwood

Everything above holds whether or not you ever speak to us. But it does reflect how we think about building AI systems for businesses. The model is the engine; the system is the vehicle. A capable model becomes useful in a business through the system built around it: the permissions it holds, the business knowledge it is allowed to use, the points where it must stop and ask, the handoff to a person, and the record it leaves behind. Those parts have to be designed, and that design is where most of the real engineering in custom software and AI applications now happens.

If you are evaluating what frontier AI could do inside your own operations, and you would rather start from a clear design than from a demo, we would be glad to talk it through. Schedule a consultation to explore what an appropriately designed AI system could look like for your organization.