K Koda Intelligence
Deep Dive
DEEP DIVE BRIEFING № 159 · 02 September 2026
FILED 02 SEPTEMBER 202624 CLAIMS CHECKED · 16 VERIFIED

Same weights, different doors
The tier is not the strategy

OpenAI's Astra is the first model it has ever rated at that capability level, and it ships in two versions: a monitored, interruptible cut for ChatGPT Plus and Pro, and an unlocked cut for vetted partners inside Daybreak Blue. Secondary reporting claiming a 100% ExploitBench score describes the restricted door, not yours. The hierarchy predates Astra: Tier 1 gets 30,000 tokens per minute on gpt-4.1, Tier 5 gets 30,000,000. Identical weights, different permission.

6 MIN READ · BY THE KODA EDITORIAL TEAM · STRATEGY · ACCESS TIERS
100%EXPLOITBENCHREPORTED, RESTRICTED TIER
30,000TIER 1 TPMGPT-4.1 PUBLISHED LIMIT
30,000,000TIER 5 TPM↑ 1,000x
LISTEN · AUDIO BRIEFINGThe conversation · ~2 min
WATCH · VISUAL NARRATIVEAnimated breakdown · ~2 min
PLAY · YOUTUBE
EXPLOITBENCH100%REPORTED, RESTRICTED TIER TIER 1 TPM30,000GPT-4.1 PUBLISHED LIMIT TIER 5 TPM30,000,000↑ 1,000x SOL OUTPUT$20.00from $30.00 SOL INPUT$4.00from $5.00 LUNA OUTPUT$1.20PER MILLION TOKENS PRIORITY TIER2x PRICEUP TO 2.5x FASTER DIFFUSION LAG12-24 MOFRONTIER TO CHEAPER MODELS

It is the first model the company has ever designated at that level. In OpenAI's own words, with the right tools and access it "can find previously unknown security flaws and develop ways to exploit them across many well-protected systems without a person guiding each step."

You cannot use Astra today. There is no calendar date. OpenAI plans to make Astra available soon, and when it arrives, ChatGPT Plus and Pro users are expected to get a monitored, interruptible version with cyber capabilities that OpenAI says will be "more limited." The unlocked version is slated to go to vetted defensive partners inside a program called Daybreak Blue, with reported participants including Cisco, Cloudflare, and Palo Alto Networks.

Same model. Two doors. One of them has your name on it, and you did not pick which.

Here is what that means for anyone building on top of these systems, why it is not actually new, and what to change in your stack before the launch you cannot schedule.

Same Weights, Different Doors

The insight is simple. Frontier capability is no longer handed out by the model card. It is handed out by the policy envelope wrapped around the model: refusal training, real-time monitors, tool restrictions, approval prompts, contract terms, hardware isolation.

ACCESS LEDGER · AUGUST 2026OPENAI PUBLISHED LIMITS · OPENAI PRICING · AXIOS

The gap between tiers is permission and throughput, not intelligence.

Tier 1 rate ceiling gpt-4.1 · tokens per minute
30,000
Tier 5 rate ceiling gpt-4.1 · same weights, 1,000x headroom
30,000,000
GPT-5.6 Sol output Per million tokens · down from $30.00
$20.00
Luna output Per million tokens · cheaper tier
$1.20

Call it Same Weights, Different Doors. The intelligence is not what varies between tiers. The permission is.

That flips how most teams evaluate. People compare benchmark scores across models and pick the winner. But a reported 100% ExploitBench result achieved under Daybreak Blue tells you nothing about what public Astra will do for you, because the public cut is reportedly set to ship with monitors that can halt a multi-step agent run and demand manual approval, even on tasks that have nothing to do with security.

The hierarchy was already there before Astra. Look at OpenAI's published rate limits for gpt-4.1: a Tier 1 account gets 30,000 tokens per minute, and a Tier 5 account gets 30,000,000. A 1,000x throughput gap on identical weights. Then layer service tiers on top, where batch runs at 0.5x price with a 24-hour window and priority runs at 2x price for up to 2.5x faster responses.

Then layer price on top of that. GPT-5.6 Sol sits at $4.00 per million input tokens and $20.00 per million output, down from $5.00 and $30.00, promotional at least through November 21, 2026. Luna sits at $0.20 and $1.20. Daybreak Blue has no published rate card at all, which is the tell: it is not a SKU, it is a relationship.

Gates Have Always Been the Business

Pull the camera back to 1996. Strong encryption was regulated as a munition in the United States. Netscape shipped two builds of the same browser: a domestic version with 128-bit ciphers and an export version crippled to 40 bits.

With the right tools and access it can find previously unknown security flaws and develop ways to exploit them across many well-protected systems without a person guiding each step.· OPENAI, ON ASTRA · 2026

Same product. Same engineers. Different door, decided by jurisdiction and paperwork rather than engineering.

That regime lasted a few years and then collapsed under its own weight. The math leaked, the controls loosened, and by the early 2000s strong crypto was a checkbox. The gate did not create a permanent class of winners. It created a temporary tax, and the firms that built their strategy on the tax got flattened when it lifted.

I think that is the correct historical frame for Astra, and it cuts against the loudest version of the tiered-access thesis. Amateurs treat a gate as a moat. Operators treat a gate as a weather condition.

Here is the asymmetry worth naming. Getting into the top tier is a small, capped upside: better throughput, earlier features, more headroom. Building your product so that only the top tier works is an uncapped downside, because you have handed a third party the right to change your unit economics and your feature set with a policy update.

There is a second wrinkle the hype misses. The teams that most want frontier cyber capability are exactly the teams who will face the heaviest monitoring, tool restrictions, and human-approval friction. Axios reported on August 7, 2026 that OpenAI slowed Astra's release specifically over these capabilities. Access and freedom of action move in opposite directions here, which is the reverse of how most people imagine a premium tier.

So what compounds? Not access. Access is rented. What compounds is your evaluation harness, your proprietary data, your workflow design, and your ability to swap the engine under the hood in an afternoon.

Only shipped work is real. The rest is a waitlist.

Whether capability gating becomes a durable industry norm or a two-year improvisation is genuinely unclear. OpenAI's own language is about risk thresholds and containment, not market structure. The company paused parts of its own development because it could not rule out Critical capability, which reads more like uncertainty than doctrine.

Permission, Not Intelligence, Is The Variable

IDENTICAL WEIGHTS
1,000x

The gate was already in the rate card.

OpenAI's published gpt-4.1 limits give a Tier 1 account 30,000 tokens per minute and a Tier 5 account 30,000,000. Service tiers stack on top, with batch at 0.5x price on a 24-hour window and priority at 2x price for up to 2.5x faster responses. Astra formalises a hierarchy that shipped years ago.

INVERTED PREMIUM
2 DOORS

The teams that want frontier cyber capability get the most friction.

Public Astra is described as monitored and interruptible, with mandatory human approval on suspicious tool calls that can halt a multi-step agent run. Access and freedom of action move in opposite directions, which is the reverse of how most buyers imagine a premium tier.

DURABLE OR NOT
12-24 MO

Restricted capability historically leaks downward.

Netscape's 128-bit domestic and 40-bit export builds in 1996 created a temporary tax, not a permanent class of winners. Frontier restrictions have tended to reappear unrestricted in cheaper or open models within roughly 12 to 24 months, with cyber-offense the one domain where governments may break the pattern.

2031

Five years out, I expect the shape to look less like a caste system and more like aviation.

Airlines do not have secret planes. They have certifications, maintenance regimes, and audited procedures, and the certification is the product. Restricted-access AI is drifting the same way. Daybreak Blue is not selling smarter tokens. It is selling a governed environment where dangerous work is legally and technically survivable.

If that holds, three things follow. Compliance-grade access becomes its own procurement line, priced by contract rather than per token. Insurance and audit firms start underwriting AI deployments the way they underwrite data centers. And a class of vendors makes real money as the paperwork layer between frontier labs and regulated buyers.

Meanwhile the floor keeps rising from below. Every capability that is restricted at the frontier this year tends to show up unrestricted in a cheaper or open model within roughly 12 to 24 months. That is the pattern from the last three years and I would bet on it continuing, with the caveat that cyber-offense may be the one domain where governments intervene hard enough to break the trend.

The strategic error in 2031 will not be landing in the wrong tier. It will be building a company whose only differentiator was a login.

Audit Your Access Tier Before You Build Another Agent

Do this in an afternoon, no CS degree required.

First, find your actual tier. Open your API dashboard and write down your rate limits: requests per minute and tokens per minute, the two ceilings that decide how many agents you can run at once. If you are on Tier 1 with 30,000 TPM, a five-agent pipeline will choke long before the model does.

Second, make your model swappable. Route every call through one function with the model name as a variable, not scattered across 40 files. Then test whether your task actually needs the flagship. Maskara AI is useful here: it runs live debates between top models and hands back the winning answer, which is a fast way to see whether $20-per-million output is buying you anything real over $1.20.

Third, design for interruption now. Public Astra is described as monitored and interruptible, with mandatory human approval on suspicious tool calls. Assume your long-running agent gets halted mid-task and build a resume path: save state after each step, log the reason for the stop, notify a human. Things will break. Test that they break loudly.

Fourth, if you are actually in security, start the paperwork. Restricted programs like Daybreak Blue are approval-gated, sold through commercial contracts with OpenAI or via AWS rather than self-serve checkout, and the review will take longer than your roadmap wants. Nobody can buy their way in with a credit card.

Fifth, keep a running log of what each tier gives you and what it costs. RemNote's mobile app lets you create spaced-repetition flashcards directly in your notes as you write, which sounds like a study tool but works well for keeping pricing and limits in your head. If you want to publish what you learn, Framesail AI takes a script through a pipeline of review steps to a near-final long-form YouTube cut. Learning in public is still the cheapest distribution there is.

One note on tool shopping: Lumi.new is live, and its official documentation lists free and Pro pricing tiers along with a feature breakdown. Read the entry before you open the landing page. An ounce of skepticism up front beats a pound of migration later.

The tier you land in is not your strategy. What you build that survives losing it is.

DOJO · BUILD THIS WEEKEND

Audit your access tier before you build another agent.

  1. Find your real ceiling. Open your API dashboard and write down requests per minute and tokens per minute. On Tier 1 with 30,000 TPM, a five-agent pipeline chokes long before the model does.
  2. Make the model a variable. Route every call through one function with the model name passed in, not scattered across 40 files. Then test whether the task actually needs the flagship: compare $20.00 per million output against $1.20 and keep the cheaper winner.
  3. Design for interruption now. Assume a monitor halts your long-running agent mid-task and demands manual approval. Save state after each step, log the stop reason, notify a human, and test that failures break loudly.
Train the full skill in The Dojo
THE BOTTOM LINE

Access is rented. Only the harness compounds.

Astra's two doors are real, and the one with your name on it was not your choice. But getting into the top tier is a capped upside of throughput and early features, while building a product that only works in the top tier is an uncapped downside, because you have handed a third party the right to rewrite your unit economics with a policy update. What compounds is your evaluation harness, your proprietary data, your workflow design, and your ability to swap engines in an afternoon. Treat the gate as a weather condition, not a moat. The strategic error will not be landing in the wrong tier; it will be shipping a company whose only differentiator was a login.

EDITORIAL RECEIPTKODA-20260902-9D569A1554B1
As of02 September 2026MethodClaim extraction, dated-evidence review, and temporal consistency gate.CorrectionsContact the Koda desk
EVIDENCE24 CLAIMS CHECKED · 16 VERIFIED · 4 FAILED
  1. 01Astra achieved a reported 100% score on ExploitBench under the Daybreak Blue program.VERIFIEDBENCHMARK
  2. 02Claim removed during the check; its text is not republished.MOSTLY FALSEMODELCUT FROM COPY
  3. 03Astra is the first model OpenAI has ever designated at the Critical level of its Preparedness Framework.VERIFIEDMODEL
  4. 04OpenAI has not announced a calendar release date for Astra, and Astra is not currently available for use.VERIFIEDFEATURE
  5. 05When Astra launches, ChatGPT Plus and Pro users will get a monitored, interruptible version with cyber capabilities that OpenAI says will be "more limited."MIXEDFEATURECORRECTED IN COPY
  6. 06The unlocked version of Astra goes to vetted defensive partners inside an OpenAI program called Daybreak Blue.MIXEDFEATURECORRECTED IN COPY
  7. 07The public version of Astra ships with monitors that can halt a multi-step agent run and demand manual approval, even on tasks unrelated to security.MOSTLY TRUEFEATURECORRECTED IN COPY
  8. 08Public Astra is described as monitored and interruptible, with mandatory human approval on suspicious tool calls.MOSTLY TRUEFEATURE
  9. 09Maskara AI runs live debates between top AI models and returns the winning answer.VERIFIEDFEATURE
  10. 10Restricted programs like OpenAI's Daybreak Blue are contractual rather than purchasable, and cannot be bought with a credit card.MIXEDFEATURECORRECTED IN COPY
  11. 11RemNote's mobile app turns notes into spaced-repetition flashcards as the user writes.MIXEDFEATURECORRECTED IN COPY
  12. 12Framesail AI takes a script through to a finished long-form YouTube cut.MOSTLY TRUEFEATURECORRECTED IN COPY
  13. 13Lumi.new is circulating with a thin directory listing, no clear pricing, and no feature breakdown.MOSTLY FALSEFEATURECORRECTED IN COPY
  14. 14Claim removed during the check; its text is not republished.MOSTLY FALSEATTRIBUTIONCUT FROM COPY
  15. 15OpenAI stated that Astra, with the right tools and access, "can find previously unknown security flaws and develop ways to exploit them across many well-protected systems without a person guiding each step."VERIFIEDATTRIBUTION
  16. 16Reported participants in OpenAI's Daybreak Blue program include Cisco, Cloudflare, and Palo Alto Networks.VERIFIEDATTRIBUTION
  17. 17Axios reported on August 7, 2026 that OpenAI slowed Astra's release specifically over its cyber capabilities.VERIFIEDATTRIBUTION
  18. 18OpenAI's batch service tier runs at 0.5x price with a 24-hour completion window.VERIFIEDPRICE
  19. 19OpenAI's priority service tier costs 2x price for up to 2.5x faster responses.MOSTLY TRUEPRICE
  20. 20GPT-5.6 Sol is priced at $4.00 per million input tokens and $20.00 per million output tokens.VERIFIEDPRICE
  21. 21GPT-5.6 Sol pricing was reduced from $5.00 per million input tokens and $30.00 per million output tokens.VERIFIEDPRICE
  22. 22The promotional GPT-5.6 Sol pricing runs at least through November 21, 2026.VERIFIEDPRICE
  23. 23Luna is priced at $0.20 per million input tokens and $1.20 per million output tokens.VERIFIEDPRICE
  24. 24OpenAI's Daybreak Blue program has no published rate card.FALSEPRICE

Every claim listed here was extracted from this article and checked against live sources before publication. The verdict is the checker's, not the writer's. Claims the check removed are counted but not republished.

Audit receipt KODA-20260902-9D569A1554B1
Filed underStrategyDeep Dive02 September 2026
Browse the Deep Dive archive

Get the morning Signal

One email per day. Unsubscribe anytime.