Source: raw/Claude_Destroyed_ChatGPT_-_The_Only_AI_You_Need_in_2026_Practical_Guide.md (auto-caption transcript; “Soul”/“Saul” normalized to Sol, “Astrea” to Astra)

Creator: AI Master (YouTube @iamAImaster) · URL: https://www.youtube.com/watch?v=OMy4hCP8uSc · Published: not in the transcript (recorded after the Opus 5.5 launch on 2026-09-22) · Sponsors: a coding-lessons app (caption reads “Codium”/“Codey”; name unconfirmed) and the creator’s own AI Master platform

AI Master gave Claude Opus 5.5 and GPT-6 Astra the same six prompts: web research, a document check, a branded website, a code-only cartoon, a video edit, and a 3D game. A side round put the cheaper models, Sonnet and GPT-6 Sol, on the website prompt. The creator scored it 5–1 to Opus but says it “felt closer than that.” Opus won the build rounds. Astra won the one round about facts you can act on and was faster almost everywhere. These are one creator’s single runs, not a benchmark.

Key Takeaways

  • Final score: Opus 5.5 five rounds, Astra one. Opus took the website, cartoon, video edit, game and cheaper-model side round. Astra took web research. The document check was a draw.
  • Astra was the safer researcher. Asked for a video-editing laptop under $1,500 with live prices and links, Opus “went wide” and wrote the better buying guide. But it quoted an old MacBook sale price and did not mention that its two Windows picks go over budget once tax is added. Astra’s prices all matched the store pages, it said its prices were before tax, and it flagged that the MacBook goes over budget with tax.
  • Both caught five of five planted errors in a travel PDF. These were a hotel starting the night after landing, one guest instead of two, no checked bags, a transfer booked at the wrong airport, and a transfer booked a day after a 00:30 return flight. Opus ranked them by how much trouble each would cause. Astra put them in a table with page numbers.
  • On build tasks Opus followed the full spec; Astra played it safe. On the hot-sauce site, Opus delivered every requested interaction, including a page-wide heat scale and a 60-second challenge. Astra built a more realistic-looking brand but changed only one section on the heat scale and skipped the challenge. In the 3D pizza-delivery game, Opus’s mini-map rotates with the rider, lamp posts count as crashes, and level three adds a storm. Astra’s camera sat too high and its levels felt the same.
  • Opus’s quality cost time. The code-only 30-second cartoon took Opus over 20 minutes against under 8 for Astra. In Claude Code and Codex, the raw-footage intro edit took Opus more than an hour because it chose the largest transcription model it could find. Astra took about 25 minutes. The creator would post Opus’s edit and called Astra’s postable.
  • The cheaper models landed close to their flagships. Sonnet built the “Kiln” site, and at the hottest heat level its glass cracks and shakes, “which even Opus didn’t do.” Sol’s site closely resembled Astra’s, down to the slogan, and like Astra it skipped reviews and the challenge. The creator’s reading is that the difference between the companies held across price tiers.
  • Per-token price and per-task cost point in different directions. Opus 5.5 is 20 per million tokens against Astra’s 50. But the video cites Artificial Analysis at about 119,000 tokens per task for Opus at max effort against about 27,000 for Astra. On rough math, a hard max-effort task “can end up costing more on Opus, not less.” In the apps, longer thinking drains the usage limit faster instead of raising a bill.
  • The creator’s split: keep Opus as the main model, and double-check anything that depends on current facts.

Round-by-round

RoundTaskSurfaceWinnerNote
1Laptop under $1,500 with live prices and store linksChat apps, web search onAstra (narrow)Opus explained more; Astra’s prices were actionable
2Find what’s wrong in a travel-booking PDF (five planted errors)Chat appsDrawBoth 5/5
3Branded hot-sauce site with scroll/mouse fire, a heat-scale slider, hover cards, a challenge sectionChat appsOpus”Saint Scoville” (Opus) vs “Bad Omen” (Astra)
430-second code-only animated short, with sound, about a pizza’s journeyChat appsOpusOpus over 20 min, Astra under 8
5Cut retakes from about a minute of raw footage, then add infographics, on-screen phrases, music and zoomsClaude Code vs Codex, separate foldersOpusOpus over an hour, Astra about 25 min
SideSame hot-sauce prompt on Sonnet vs GPT-6 SolChat appsSonnet”Kiln” vs a near-copy of Astra’s site
6Playable 3D night-city pizza-delivery game (“Midnight Slice”, both models picked the name)Chat appsOpusScope capped at a few blocks, 3–5 minutes of play

Benchmarks quoted in the video

Most of these figures match what the wiki already records in Claude Opus 5.5 and GPT-6. The new items are Astra’s token count and the hallucination test.

  • Artificial Analysis Intelligence Index: Opus 5.5 58, Astra 53. The video does not name the index version.
  • Artificial Analysis hallucination test: Astra is less likely to make up an answer when it should say it doesn’t know. The rate is 51% for Astra against 59% for Opus (lower is better, as the video frames it).
  • Tokens per task at max effort: about 119,000 for Opus and about 27,000 for Astra.
  • Anthropic’s table: agentic coding 66.4% vs 57.9%, very hard expert questions 67.7% vs 57.2% (Opus leads). Business automation 40 vs 41.4 and long scientific calculations 58.7% vs 64.6% (Astra leads). These match the Terminal-Bench 4.0, Humanity’s Last Exam, AutomationBench and Terminal-Bench Science rows already in the wiki.

Try It

  1. Route by task type. Use Opus for builds where following every item in a long spec matters (sites, games, edited video). For buying decisions and anything that depends on today’s prices or stock, use Astra or verify Opus’s links by hand. This is the creator’s own split.
  2. Make research prompts state the constraint you care about. Round 1 turned on tax and current stock. Name both in the prompt (“prices after tax, in stock today, link to the listing”) rather than relying on the model to raise them.
  3. Budget time, not just tokens, for Opus build jobs. Plan on 2–3x Astra’s wall-clock time for code-only animation and local video editing, per these runs.
  4. Try Sonnet first on build prompts. In this test the cheaper Claude model got “very close” to Opus on a demanding front-end prompt.
  5. Check effort before you compare cost. At max effort Opus’s token use can outweigh its lower per-token price. See Which Effort Level to Use.

Open Questions

  • Which effort levels were used? The video quotes max-effort token counts from Artificial Analysis, but it does not say which effort each model ran at in the six rounds.
  • Which Sonnet? The side round says “Claude Sonnet.” The recording date isn’t given, so it could be Sonnet 5 or Sonnet 5.5 (launched 2026-09-28). Likewise, “GPT-6 Sol” could be GPT-6 Sol or GPT-6.1 Sol.
  • Hallucination metric direction. The video says Astra’s 51% means it is less likely to make up an answer than Opus’s 59%. The exact AA metric and version aren’t named.
  • Sponsor name. The coding-app sponsor’s name is garbled in the captions.