AI news videos

Asana's 76x Figure Is Two Separate Gains Multiplied

OpenAI published a customer story on October 9, 2026 saying Asana made its browser agent 76x cheaper and 5x faster in tests, using GPT-6 Astra in Codex to run the experiments and GPT-6.1 Sol to run the optimized workflow. Everything below comes from OpenAI's own announcement.

Watch on YouTube

Transcript

Sources

Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:

• Scene by scene evidence ledger, quoted from the OpenAI page itself, which is the only fact source used. THE HONESTY SPINE FIRST, because it governs every line: what is genuinely new on October 9 2026 is that OPENAI PUBLISHED THIS PAGE. No product launched. The engineering work described is already in production, because the page itself says "Asana has released the changes to browser navigation in StackAI and is developing tools to make similar experiments easier to repeat." Accordingly nothing in this script says or implies "launched today", "just launched", "announced today", "new today" or "rolling out" about anything Asana or OpenAI makes; scene 1 narrates the literal equivalent, "nothing launched today", and the scene 1 card adds the publisher's own sentence about the changes having been released as the proof. Our old disclosure sentence "There is no product launch here." is deliberately NOT reused anywhere in this file, because it has sent our transcriber into a repetition loop on two prior runs. HOOK BACKING (OpenAI says Asana cut model costs seventy six times, and the arithmetic matters): the first half is the page's own headline, "Asana cuts model costs 76x in browser tests with GPT-6.1 Sol", and its first stat block, "76x / Lower estimated model costs with the optimized GPT-6.1 Sol workflow". The attribution verb is "says", not "did", because the figure is the publisher's claim about its own product and is an estimate in the page's own word. The second sentence, "The arithmetic matters.", is a statement of the Short's editorial subject and asserts no fact; it is cashed out in scene 4, where the 76x is decomposed into the page's own component numbers. The hook is exactly two sentences, because our hook card renders two beats and breaks with only one, and it is 12 words, inside the 20 word hook ceiling. SCENE 1 (OpenAI published this customer story about tools it sells, and nothing launched today): the publisher is OpenAI, the byline is "OpenAI" with no named human author, the metadata block reads "Company size: Enterprise", "Region: North America", "Industry: Technology", "Products: Codex", "Contact sales" appears twice, "Customer Stories" is a standing nav entry on openai.com, and the page closes with "Join the new era of work" above "More than 1 million businesses around the world are achieving meaningful results with OpenAI." OpenAI sells Codex and it sells the model the page calls "OpenAI's model" in its own comparison table, so "tools it sells" is literally true of both named products. The nothing launched half is the honesty spine stated above and rests on "Asana has released the changes to browser navigation in StackAI". The "new era of work" pitch and the 1 million businesses line are shown on the card strictly as marketing furniture and are never spoken. SCENE 2 (an engineer directed GPT six Astra to investigate Asana's browser agent, which kept resending page history at full price): the verb is the page's own, verbatim "He directed GPT-6 Astra in Codex to investigate the agent, test improvements, and compare the results." The finding is verbatim too: "GPT-6 Astra discovered that the agent cached its fixed instructions and tool definitions, but not the growing history of page text and screenshots it gathered, so every request resent that history at full price." Narration says "an engineer" rather than naming the man or his title, because titles like CTO and the post nominal PhD are confined to silent cards by our pronunciation rules; the card carries "Asana's StackAI CTO, Frank Hidalgo, PhD" in full. Narration says "Asana's browser agent" and never says the product name StackAI, which is likewise confined to the card. The three candidate fixes are a three item list and are therefore on the silent card only, verbatim, never spoken. SCENE 3 (Astra completed one hundred forty four runs in a week, not an estimated two months): the run count is verbatim, "Asana's 144-run study tested GPT-6.1 Sol and three other frontier models, called here Models A, B and C", and the conductor is verbatim, "Astra conducted the full study". The week against months contrast is the page's own section heading, verbatim "From an estimated two months of research to one week with GPT-6 Astra", and the engineer's own quote, "This would have taken me one to two months by hand. With GPT-6 Astra in Codex, it took about a week". Narration says "not an estimated two months" and keeps the word "estimated" precisely because this is the engineer's estimate of a hypothetical he never ran, not a measured control arm, and the card states that limitation explicitly. The banned idiom "by hand" appears nowhere in narration; it survives only inside the quotation marks on the scene 3 card, clearly marked as the source's own words. The study design detail, which is six caching and screenshot policies, two history budgets of 120,000 and 480,000 characters, three repeats per cell, four models and one fixed task of collecting six fields for each of 32 books, is a multi item list and is on the silent card only. SCENE 4 (the workflow fix was worth twenty nine times on a rival model, and switching to OpenAI's model multiplied that by two point six): this is the editorial heart and it is pure arithmetic from the page's own sentence, verbatim "For Model B, the optimization reduced estimated model cost from at least $36.21 (some original runs reached the step limit before they finished) to $1.24 per run, a reduction of 29x. The optimized workflow on GPT-6.1 Sol was 2.6x cheaper still, at $0.47." So the 76x headline is a 29x workflow change measured on a competitor model, multiplied by a 2.6x model switch to the publisher's own model, and the Short says so in those terms. Narration says "a rival model" rather than Model B because the page anonymises every competitor, and says "OpenAI's model" rather than spelling out the model name, both to keep the sentence honest about who benefits and to avoid a brand name glued to digits in speech. SCENE 5 (some original runs never finished, so that baseline is a minimum, no rival is named, and we confirmed nothing independently): the first clause is verbatim parenthetical from the page, "some original runs reached the step limit before they finished", and the lower bound reading is the page's own chart footnote, verbatim ">=: the baseline includes capped runs, so its mean is a lower bound" and ">=: mean includes a capped or unfinished run, so the true value is at least this large". Narration says "a minimum" as the plain English equivalent of "a lower bound". The anonymisation clause rests on the model table, which identifies competitors only as "Model A", "Model B" and "Model C" with descriptions such as "from another frontier lab, released Fall 2025" and never names a lab or a model, so no reader can check which model was beaten. The no independent confirmation clause rests on our own corroboration status: a single first party source, a 403 on the live URL so that all text came from our same day snapshot, 0 of 5 video transcripts retrieved, and 16 timeline probes that were unrelated noise. The card adds that both named voices, Frank Hidalgo and Arnab Bose, work for Asana, that no analyst, journalist, regulator, competitor or independent researcher is quoted, and that no independent audit of any figure is claimed anywhere on the page. FAIRNESS NOTE, deliberate: this study is better evidence than most vendor pages carry, and the Short credits it twice, once by narrating the specific run count in scene 3 and once by putting the full design, the fixed task and the published study on the cards. The criticism is aimed only at the headline multiplier and the anonymised comparison, never at whether the engineering happened. EXCLUDED ON PURPOSE, and why. The 5x speed figure, the $0.47 average cost, the 4x cost reduction on GPT-6.1 Sol alone, the 89% cache share, the 22.5 minute original runtime and the roughly four minute optimised runtime are all EXCLUDED FROM NARRATION to protect the word budget and, more importantly, the no repeated figure rule; every one of them is on the silent cards verbatim with digits. The genuinely good finding that a larger history budget raised runs producing an answer from 3 of 18 to 18 of 18 is EXCLUDED FROM NARRATION for the same budget reason and sits on the scene 3 card under the heading ONE GENUINELY GOOD FINDING, so it is still on screen. Any claim about Asana's finances, its customer count, its contract with OpenAI, or any commercial relationship beyond what this page states is EXCLUDED FOR LACK OF CORROBORATION: the page does not mention one, we hold zero corroborating transcripts, and our rule requires two independent confirmations for a claim absent from the article, so no such claim is asserted, implied or hinted anywhere in this file. The 16 timeline probe titles were unrelated tutorials and appear nowhere in narration or on screen text, not even by implication. The 0 of 5 YouTube transcript result is recorded as a tooling fault, not as evidence of anything. The three "Keep reading" customer stories about Oracle, 1Password and loveholidays are separate posts and were not used as evidence for any claim here. The forward looking material was kept out of narration as intent only, specifically "Asana is now using GPT-6 Astra in Codex to test product features before release" and "Over time, the team plans to incorporate this testing into the platform's evaluations". All site chrome, nav lists and footer policy links were ignored and contributed nothing. GATE NOTES. WORD COUNT: 116 narrated words in total, inside the required 112 to 118 band, counted as hook 12, scene 1 13, scene 2 19, scene 3 15, scene 4 22, scene 5 20 and endcard 15. PER SCENE: 13, 19, 15, 22 and 20, every one of them under the renderer's undocumented 30 word per scene cap, with the largest at 22. SENTENCES: 11 in total, under the 13 sentence ceiling, counted as hook 2, scene 1 1, scene 2 1, scene 3 1, scene 4 2, scene 5 2 and endcard 2; three of the five scenes are a single sentence on purpose, because our voice slows down on short sentences and a lower sentence count buys duration headroom. COMMAS: deliberately sparse, with at most one comma per narrated sentence, because a comma heavy sentence reads slower and has overrun 60 seconds before. LISTS: narration carries no multi item list anywhere. The three candidate fixes, the study's design parameters, the four models, the three stat blocks, the page's four figures of speech and the two named Asana voices are all on silent cards only. DASHES: no em dash, no en dash and no double hyphen appears in any field of either file. The source's em dashes before quote attributions were recast with parentheses, and the only ASCII hyphens that survive are inside quoted source strings on silent cards, such as "144-run study", "best-performing policy", "GPT-6.1 Sol" and "GPT-6 Astra". QUOTES: every curly quotation mark and curly apostrophe from the source was converted to straight ASCII. IDIOMS: the source is dense with figures of speech and every one is barred from narration, specifically "brings human and agent teams to life", "what teams of humans and agents look like in practice", "set the direction", "add up", "Shipping speed is no longer the bottleneck", "human attention is", "a fleet of agents", "every engineer is a PM", "the new era of work", "by hand", "cutting back" and "move quickly". Narration states the literal thing plainly instead, and every one of those phrases appears in this file only inside quotation marks on a silent card, clearly marked as the source's own words. The source's own verb "directed" was reused in scene 2 precisely so that narration did not need a figurative substitute. MODEL NAMES: narration spells the only model it names as "GPT six Astra" on first mention in scene 2 and then uses the bare word "Astra" in scene 3, both of which are safe in our pronunciation vocabulary. The second model is never named in narration at all, and is referred to as "OpenAI's model" in scene 4, which is exactly how the page's own comparison table describes it; this deliberately avoids speaking "six point one" next to "two point six" in the same scene. No hyphenated or digit glued form such as GPT-6.1 or GPT-6 appears in any narration field; those forms appear only on the silent cards. CONFINED TERMS: StackAI, Command, SQL, PhD, CPO, CTO, PM, every URL and the word agentic appear nowhere in narration and only on silent cards. Narration says "Asana's browser agent" for the product and "an engineer" for the person. NUMBERS: every narrated number is written as words, specifically "seventy six times", "one hundred forty four", "two months", "twenty nine times" and "two point six". All digits, including 76x, 5x, $0.47, $1.24, $36.21, $1.97, 29x, 2.6x, 4x, 89%, 22.5 minutes, 120,000 and 480,000 characters, 32 books, 3 of 18 and 18 of 18, live on the silent cards only. FIGURE REPETITION: no figure is spoken twice anywhere in the script. "Seventy six times" is spoken once, in the hook, and scene 4 deliberately decomposes it without re speaking it. "One hundred forty four" is spoken once, in scene 3. "Twenty nine times" and "two point six" are spoken once each, in scene 4, and are the only two multipliers in that scene. The one near collision we accepted and checked is the bare word "two" inside "two months" in scene 3 and inside "two point six" in scene 4; these are two different figures in two different scenes roughly eight seconds apart, with different following words, so they are not a repeated figure under the rule and do not present the comma and zero runaway pattern that has looped our transcriber on money figures. DATES: no date is spoken anywhere, in narration or the endcard; every date appears as digits on silent cards only, as Oct 9 2026, Oct 8 2026, Sep 8 2026, Aug 26 2026, Fall 2025, Summer 2026 and Fall 2026. QA PHRASES: the exact phrase "like and subscribe" is present in the endcard narration, and no countdown or deadline language appears anywhere, with the word that trips our countdown gate absent from every field. SUBLINES: every cards.json subline is between 180 and 250 characters, because the subline is rendered on screen and overflows the frame past about 250 characters. STRUCTURE: exactly 5 scenes, each durationS 8 so the QA gate's sum is 40, inside its 30 to 50 window; those integers are cosmetic planning values the renderer never reads. NO timeline array is included, deliberately, so our date resolver keeps the same day storyDate of 2026-10-09 rather than preferring an earlier timeline date and wrongly aging this same day publication into the catalog tier.(HOOK AND FULL-SCRIPT BACKING. First party OpenAI customer story, headline "Asana cuts model costs 76x in browser tests with GPT-6.1 Sol", deck "Using GPT-6 Astra in Codex, Asana made its browser agent 76x cheaper and 5x faster in tests to offer customers more capable models.", byline "OpenAI" with no named human author, dated on the page October 9 2026. The page carries the standard customer story furniture: a "Contact sales" call to action twice, the metadata block "Company size: Enterprise", "Region: North America", "Industry: Technology", "Products: Codex", the standing nav entry "Customer Stories" on the same host, and the closing pitch "Join the new era of work" above "More than 1 million businesses around the world are achieving meaningful results with OpenAI." This is therefore a vendor marketing asset about a customer of the vendor, published by the company that sells both the tool and the model being praised, and the Short says so on camera in scene 1, in the viewer's first seconds. The live URL returns HTTP 403 to our fetcher, which is normal and expected behaviour for openai.com, so the only text we hold is our watcher's snapshot, captured at 2026-10-09T18:35:23.290Z, the same day the page is dated. That snapshot IS the primary source for this Short and is not a degraded fallback: it is the publisher's own first party page on the publisher's own host, not a link out stub, aggregator copy or syndication. The page's metadata line "Region: North America" and the fact that Asana is a United States company mean no regional caution applies. CORROBORATION STATUS: ZERO. The YouTube transcript pull returned 0 of 5 videos because the installed yt-dlp on this host cannot fetch captions, which is a known tooling fault and not a content problem. The 16 dated timeline probe titles came back as completely unrelated content, specifically a Microsoft Loop explainer, a Chrome cookies Short, a WIX code snippet tutorial, a Google Play Store fix, Google Tasks tips, a Pipedrive discount clip, two Copilot and Outlook tutorials, three Claude tutorials and a build your first agent tutorial; they corroborate nothing, they were discarded entirely, and they are asserted, implied and hinted nowhere in narration or on screen text. This story is single sourced and the final scene states that on camera, in the narrated words "we confirmed nothing independently", with the single source position spelled out in full on the scene 5 dossier card. The three "Keep reading" links at the foot of the page, "How Oracle turns days of work into minutes with ChatGPT and Codex" (Oct 8, 2026), "1Password increases engineering productivity 21% with Codex" (Sep 8, 2026) and "How loveholidays is making everyone a builder with Codex" (Aug 26, 2026), are separate customer stories about other companies and were not used as evidence for any claim here. No prompt injection or instruction like content was found anywhere in the source text; it was treated throughout as untrusted data informing facts only, never as instructions.)