Video summary

It's Here.

Main summary

Key takeaways

Technology

Summary of technological concepts & key product/benchmark takeaways (GPT-6 Astra)

Major model release & positioning

  • The speaker frames the announcement as the long-awaited “next-gen” model: GPT-6 Astra.
  • Astra is presented as a generational leap over the prior line (especially “Soul”/previous GPT-5.x variants), with improvements across:
    • Code
    • Agentic computer use
    • 3D / creative tooling
  • The speaker notes there are not yet other GPT-6 siblings (e.g., “Soul and Luna and Terra equivalents”)—Astra is the primary one covered.

Scope of evaluation the video claims to cover

The video says it will address:

  • Availability
  • Cost
  • Security / safety
  • How it works day-to-day
  • Demos and projects built with the model

It also highlights:

  • Temporary access limitations
  • “weird restrictions” during rollout

Cost & pricing structure (and what changes the real cost)

Base token pricing (comparable to previous model)

  • Astra is cited at roughly:
    • ~$10/m input tokens
    • $50/m output tokens
  • The speaker claims this is nearly identical to Fable-class pricing.

Fast mode

  • Astra includes fast mode (contrasted with Anthropic’s Fable, which is said to only have fast behavior in their “Opus” tier).
  • Fast mode details:
    • Up to ~2× speed
    • ~2× price

Token efficiency dominates real-world cost

  • The speaker claims Astra is “insanely efficient”.
  • In some cases, it’s described as even cheaper than Soul for comparable tasks.

Caching / context read pricing difference

  • Astra is said to not have the same caching read discount the speaker claims Fable has.
  • Claimed pricing:
    • Astra cached reads: $1 per million cached token reads
    • Fable cached reads: $0.25 (speaker notes a 75% reduction vs something else)
  • The speaker still concludes Astra is cheaper overall because efficiency outweighs the caching disadvantage.

Context window & multiplicative pricing behavior

  • Astra supports up to ~1M token context, but availability is nuanced:
    • Not available by default in “CodeX” (cloud IDE mentioned).
    • Likely staged availability (speaker suggests ~370k default plan).
  • Cost scaling when context exceeds ~272k:
    • Input tokens cost ~2×
    • Output tokens cost ~1.5×
  • The speaker also mentions an exception/mitigation in CodeX, implying costs may not multiply exactly as raw pricing would suggest.

Availability / rollout / access constraints (important caveat)

Rolling out to limited orgs first

  • Astra is rolled out initially to a small subset of organizations.
  • Expansion is described as happening over days, eventually reaching:
    • ChatGPT Plus / Pro
    • business / enterprise

Cloud IDE limitation frustration

  • The speaker is upset that Astra was not available to them in CodeX at the time of filming.

Enterprise privacy

  • Claims Zero Data Retention (ZDR) is provided.

Where it’s available

  • Available via:
    • OpenAI API
    • Amazon Bedrock
  • The speaker notes there is no mention of Azure, implying Azure integration may be missing (consistent with a post-company-split assumption).

Temporary access “reset” workaround

  • The speaker mentions a tweeted workaround:
    • Daily “banked reset” credits for CodeX users who don’t yet have Astra access.
  • OpenAI is said to have confirmed this is temporary.
  • The speaker speculates it could be tied to compute/enterprise timing or partner rollout.

Benchmark highlights (the speaker’s “Astra wins” narrative)

The video repeatedly claims Astra scores are:

  • Higher across multiple areas
  • Often cheaper for similar tasks due to efficiency
  • Sometimes shows oddities like:
    • “X high worse than high”
    • “max sometimes not best”

Examples of benchmark categories mentioned

Terminal Bench (computer / terminal tasks)

  • Astra is said to beat other models (including Fable 5.1) at similar or lower cost.
  • The speaker provides example numbers (with large percentage gaps vs Claude/Fable).

Science / Math

  • Science benches: Astra is described as “slaughtering” rivals and being best + cheapest.
  • Frontier Math:
    • Fable best ~87.8%
    • Astra reaches ~100% at medium+ levels

Agentic / tool-use style bench (Arc AGI mention)

  • The described bench is punitive about:
    • step count
    • tool-call efficiency
  • Despite that, Astra is claimed to achieve near 99.9%.

Cybersecurity

  • Exploit Bench:
    • Astra claimed to hit 100% at “low”
    • The speaker notes a potential table oddity (where “high” might be cheaper than “low”)
  • Compared against prior best scores from Soul on max.

Alignment / safety stress tests

  • Astra is said to be “most aligned” in sensitive areas.
  • Example: a computer-use safety stress test where lower misalignment is better.

Important nuance about evaluation

  • The speaker critiques the “Artificial Analysis” bench:
    • Astra ties with some models there but doesn’t match its usual dominance.
    • The speaker argues the bench suite may be outdated and not reflecting modern behavior well.
    • The speaker cites feedback from the bench founder about overhauling it.

Computer use / agents: the strongest “real-world capability” emphasis

Operating a computer

  • Astra is framed as dramatically better at operating a computer:
    • Prior agents (especially Claude/Fable-type) are described as feeling “stone age” for computer use.
    • Astra is claimed to use mouse/keyboard confidently.

Speed advantage (example)

  • Example time comparison:
    • Astra: OSWorld 2.0 in ~23 minutes at ~71.6%
    • Soul best: ~1h 15m at ~65.7%
  • The speaker claims Astra is both faster and higher-scoring.

Benchmarked agent performance

  • Mentions improvements and comparisons across:
    • agents from a “last exam”
    • screenshot pro
    • OSWorld
  • Also notes cost differences.

Real app demos

  • Demonstrates Astra driving Excel via cursor/mouse + typing.
  • Claim: it “thinks first” (startup delay), then executes quickly once committed.

3D modeling and creative tooling

Blender + CAD + game pipeline claims

Astra is presented as very strong at:

  • 3D scene generation
  • tool-aware 3D workflows
  • translating outputs into Unreal Engine 5

BenchCAD (Python CAD benchmark)

  • Astra is claimed near ~100% peak accuracy.
  • The speaker says it outperforms Fable’s best at lower cost.

Demos described as “absurdly good”

  • Includes claims of:
    • walkable scenes
    • detailed geometry
    • shaders / lighting
    • animations
  • Speaker claims it produced “fish/gameplay-like loops” and Unreal-ready environments.

But not perfect

  • One weakness described: when asked to build/animate a game controller/UI, Astra’s control handling may be “janky” compared to Fable 5.1.

Interaction quality improvements (how it “talks” and plans in tools)

  • Better instruction interpretation
    • Astra allegedly chooses better defaults when instructions are ambiguous.
  • Focused questions
    • Can ask a question without derailing the workflow (described as asynchronous behavior).
    • Improved CodeX behavior: asks while continuing other dependent work.
  • Staying oriented as tasks evolve
    • Reduced tendency to treat steering messages as “new goals,” helping it keep constraints.

Code / review workflow claims (agentic software engineering)

PRs, tests, and scope control

  • Contrasted Astra vs Soul:
    • Soul is described as leaving “messes behind,” overengineering PRs, writing unnecessary tests, and going too far.
    • Astra is framed as more restrained and better at understanding scope of change.

Example: using Astra to fix and respond to review comments

  • The speaker describes Astra creating PR changes and responding to automated review comments.
  • They also mention failure modes later in the “rough edges” section.

Real-world projects built with Astra (tutorial-style “what it can do”)

The speaker lists concrete product-like examples (not formal tutorials):

  • Spotify clone
    • Converts a blog-based playlist/music writeup site into a working player with improved navigation/resume behavior.
  • Plex-like media server/client
    • Builds streaming for NAS/local network and via Tailscale.
    • Claimed to become the speaker’s primary media player after multiple iterations.
  • “LakeBed” cloud project performance overhaul
    • Astra audits performance and stress tests the service.
    • Reported improvements:
      • latency reduced from up to ~800ms to <30ms
      • P95 shaved by up to ~98%
    • Speaker claims changes were safe (no regression/security issues), emphasizing:
      • sub-agent / swarm coordination
      • merging many PRs
  • YouTube Shorts editing attempt
    • Speaker reports disappointment: Astra is said not to edit video well (clip cuts/layout are described as “cringe”).
  • 3D game demos
    • Mentions community creators and examples like:
      • Minecraft clone
      • open-world FPS
      • small planet walkable games
      • Fall Guys clone
    • Claim: model can build a detailed walkable Manhattan in Unreal over a week.

Rough edges / limitations (explicitly called out)

  • Access restrictions prevent replication
    • Some workflows can’t be shown because Astra isn’t enabled for the speaker in CodeX at filming time.
  • Agent reliability issues
    • Even with a newer snapshot, “bad behaviors still occur.”
  • Problematic PR-monitoring behavior
    • Example:
      • Astra finds a fix quickly and opens a PR.
      • Later, when addressing additional automated review comments:
        • it doesn’t always apply changes immediately
        • monitoring sometimes stops before all review comments arrive
        • it may say “yes” it will address comments but fails to push changes
    • The speaker says OpenAI later adjusted model/harness/system prompt layers, but issues can recur.
  • Overengineering / loops
    • Less than Soul, but Astra can still get “trapped” responding to many review comments.

Sponsored tool mentioned (workflow support)

Code Rabbit

  • Sponsored tool integrated with IDE + CLI + GitHub
  • Includes “Change Stack” for timeline + chunked PR review
  • Mentions an agent that can read:
    • Slack
    • Linear / Jira
    • emails
    • plus correlating alerts (e.g., DataDog) to PRs and follow-ups
  • URL shown: soy.link/codabbit

Main speakers / sources

  • Main speaker: the video author/reviewer (single primary narrator; no name given in subtitles)
  • Referenced external sources / creators / figures:
    • Anthropic (comparator: Claude / Fable)
    • Arc AGI (bench creator referenced)
    • Code Rabbit (sponsor; features referenced)
    • OpenAI (model provider; rollout/policy described)
    • Demo builders mentioned:
      • Flavio
      • Peter
      • Matthew Burman
      • Matt Schumer
      • Max Weinbach
      • (plus others in demo context)
  • Direct social/source callouts:
    • Hebo (tweet about CodeX banked resets)
    • Dar (witchai.dev/design comparisons mentioned)

Original video