Claude Opus 5.5 vs. Opus 5 vs. Fable 5.1: Ich habe nachgerechnet

Product reviewed

A comparison of three AI model options (from the same provider/interface) for code/task generation and problem-solving:

• Opus 5.5
• Opus 5 (predecessor)
• Fable 5.1

Tests were run with the same prompts and (for some parts) the same “medium” difficulty level, using Cloud Code / interface token costs (no tools, no rework).

---

Key features / behavior observed (from the video)

• Ability to generate a single-file browser game (lunar lander autopilot, fuel limits, scoring, explosion vs clean landing, etc.).
• Ability to:
  • Proofread/check and correct a structured “bathroom quote” document containing multiple intentional errors.
  • Solve a logic/puzzle scheduling problem with constrained clues (exactly one solution expected).
• The models expose an “Effort” setting affecting speed and cost:
  • For Opus 5: low / medium (default) / high / extra high / maximum
  • Opus 5.5 shows warnings at the highest effort (usage limit usage increases).

---

Pricing / token-cost findings (numerical)

Task 1: Moon landing game (browser game in one file)

• Opus 5 average cost: $0.89
• Fable 5.1 average cost: $1.46
• Opus 5.5 average cost: $1.90 (stated as more than 2× Opus 5)

Notes on “effort/difficulty”:

• Claims that Opus 5.5 is often ~40% cheaper than Opus 5 for typical work/settings.
• But in this specific test at medium, Opus 5.5 was >2× more expensive.
• With a different setting, Opus 5.5 could complete the same task for $0.41, which is less than half of Opus 5 “Medium” (and the landing worked reliably).

Token usage (Task 1 / general measured behavior)

• Opus 5: ~35,000 total tokens
• Opus 5: writes ~95,000 tokens mentioned for “thinkers” (with ~65,000 thinkers)
• Fable 5.1: ~29,000 tokens
• Opus 5.5: “takes much longer to think” and writes the longest program on average (highest effort/compute behavior).

---

Task 2: Bathroom quote error detection/correction

• All three models found 8 errors in every run with no false alarms.
• Time to generate the error list:
  • Opus 5: ~24 seconds
  • Opus 5.5: ~36 cents/test (time not clearly tied to a single value here beyond Opus 5 and Fable 5.1 times)
  • Fable 5.1: ~31 seconds
• Estimated cost per test:
  • Opus 5: ~$0.06
  • Opus 5.5: ~$0.09
  • Fable 5.1: ~$0.16
• Corrected final amount: final total was €337 less than the original offer.

---

Task 3: Logic puzzle scheduling

• All three models found a valid plan every time; with clues there is exactly one solution.
• Time and cost:
  • Opus 5.5: ~20 seconds, ~$0.05
  • Opus 5: ~50 seconds, ~$0.12
  • Fable 5.1: ~~30 seconds, ~$0.14
• Conclusion for Task 3: Opus 5.5 fastest and cheapest with the same correct result.

---

Interface price per token (Opus 5.5)

• Opus 5.5 via interface: $4 input and $20 output per million tokens (~20% less than Opus 5).
• Also mentioned: Artificial Analysis index ranks Opus 5 first place overall (independent measurement).

---

User experience / reliability observations

Task 1: Game landing reliability

• The game outputs were visually different across models.
• Landing success consistency:
  • Opus 5: landed in 5 out of 9 runs (at medium settings described in the landing-frequency comparison)
  • Fable 5.1: landed in 7 out of 9
  • Opus 5.5: described as “Perfect landing” in the main comparison
    • A later claim says Opus 5.5 and Opus 5 had the same results for selection/puzzle tasks, but the game landing-frequency is specifically highlighted as:
      • Fable did better than Opus 5 in that test
      • Another statement: at “standard high” level Fable landed at 2.70 on all three starts (interpreted as consistent performance for that trial)

Notable bugs / output issues

• Opus 5 (one run): game wouldn’t start (bug prevented the “ferry” from appearing).
• Another Opus 5 output: lunar surface rendered as a “comb” of jagged edges.
• Opus 5 took the longest for at least one generated version.

Performance / speed (high level sentiment)

• Opus 5.5 generally improved speed/cost on the puzzle task, but game task behavior depended heavily on effort level.
• Fable 5.1 can be slower and more expensive in general, but is positioned as strong for very difficult thinking / long agent work.

---

Pros (as evidenced in the video)

Opus 5.5

• Strong cost/speed for puzzle solving (Task 3): fastest and cheapest with correct results.
• Better token/input pricing vs Opus 5 (interface: ~20% less input/output per million tokens stated).
• “Effort” system allows selecting lower effort to reduce cost while still getting correct results in some scenarios.

Opus 5

• Best overall efficiency in parts of testing:
  • Task 2 cheaper than the others
  • For medium, stated as fastest/cheapest among the three for offer/puzzle-style work
• Exposes thinking process in the interface (useful for verification UX).

Fable 5.1

• More reliable than Opus 5 in game landing frequency during the described medium test (7 vs 5).
• Designed for very difficult thinking / longer agent work.

---

Cons (as evidenced in the video)

Opus 5.5

• Can be much more expensive than Opus 5 at the same medium effort in the game task (>$2× in the pricing summary).
• Higher “effort” can consume usage limits faster (explicit warning behavior at extreme effort levels).

Opus 5

• Lower game reliability (some runs failed to start; visual/geometry issues; landing frequency lower than others in that test).

Fable 5.1

• Generally more expensive; for Task 2 it was the most costly per test (~$0.16).
• In at least one comparison, it was not better on speed for the quote-check task (31s vs Opus 5 at 24s).
• Subscription/limits behavior is restrictive:
  • Max tier only allows using half the weekly limit for Fable models.
  • Pro tier may exclude Fable from the limit (may require paying separately via usage credit).

---

Comparisons / deciding guidance from the video

The tester’s practical recommendation:

• Use Opus 5 as the default for checking/calculating and as the generally best baseline.
• Use Opus 5 on low when you need speed/cheaper outputs and medium is sufficient.
• For large construction projects (like full game generation):
  • start with low
  • move to medium only if needed
  • reserve maximum for special cases
• Use Fable 5.1 only if Opus 5 fails, since it’s more expensive and “meant” for harder, longer reasoning.

---

Overall verdict (concise)

• Best default for cost-effective correctness (quote + puzzle style): Opus 5 (medium/low depending on need).
• Best for puzzle-type tasks where speed/cost matters: Opus 5.5 (fastest/cheapest in Task 3).
• Best when you need stronger/longer thinking or more reliable game landings (in that medium test): Fable 5.1, but it costs more and is harder to justify unless Opus 5 underperforms.

Strictly by the video’s test results: Opus 5 as the primary workhorse, Opus 5.5 as the upgrade for faster puzzle-style tasks, and Fable 5.1 as a fallback for harder/longer reasoning or specific reliability needs.

---

Speakers / perspectives

• Single main speaker/tester narrating the measurements and recommendations.
• No separate speakers are explicitly quoted beyond the narrator’s references to other entities (e.g., “Artificial Analysis,” “Enthopic”).