The Vibe Father
Model Review
Claude Opus 4.8 is Anthropic's senior workhorse, released May 28, 2026 — eleven days before Fable 5 stole its crown. That timing shaped its story unfairly. Judged on its own, Opus 4.8 is an exceptional coding model: 88.6 on SWE-bench Verified, a 1M-token context window, and half the price of the flagship. We have been running it hard since release, and our honest read is that for most serious engineering work, this is the model to actually buy.
Pricing is $5 per million input tokens and $25 per million output, streaming at 60 tokens per second. It is not fast — Sonnet 5 streams half again quicker — but it is deliberate in a way that shows up in the work: fewer hallucinated APIs, fewer confident wrong turns, better long-horizon coherence across big diffs.
The numbers
Here is where Opus 4.8 sits against its neighbors on our live leaderboard at /benchmarks, where the Vibe Coding Index weights SWE-bench Verified 40%, Terminal-Bench 30%, and LiveCodeBench 30%.
| Model | SWE-bench Verified | Terminal-Bench | LiveCodeBench | Price (in/out per M) |
|---|---|---|---|---|
 Claude Opus 4.8 | 88.6 | 78.9 | 87.8 | $5 / $25 |
| Claude Fable 5 | 95.0 | 83.1 | 89.8 | $10 / $50 |
| Claude Sonnet 5 | 85.2 | not yet published | 82.4 | $3 / $15 |
 GPT-5.5 | 80.6 | 83.4 | 85.3 | varies by tier |
Read the table honestly and the positioning is clear. Fable 5 beats Opus everywhere, by 6.4 points on SWE-bench Verified — but costs exactly double. Sonnet 5 trails Opus by 3.4 points on SWE-bench and 5.4 on LiveCodeBench at 60% of the price. GPT-5.5 beats it on Terminal-Bench (83.4 vs 78.9, and GPT-5.5 holds the top Terminal-Bench score on our whole board) but loses by 8 points on SWE-bench. Opus 4.8 wins no single event outright. It just refuses to be bad at anything.
Day to day, that all-rounder profile translates into something we value more than any single score: predictability. Opus 4.8 almost never produces the catastrophic miss — the confidently wrong architecture, the refactor that silently drops an edge case. Its failures are boring and legible: it stops, says what it is unsure about, and asks. For long agent sessions, that temperament matters as much as raw capability, because a model that fails loudly costs you a retry while a model that fails quietly costs you a debugging day. The 1M context also earns its keep on real repositories — we regularly hand it entire feature areas plus tests plus docs and it keeps the threads straight deep into the session.
The price math
Our reference heavy month is 50M input tokens and 10M output tokens — realistic for someone running coding agents daily.
Opus 4.8: 50 × $5 + 10 × $25 = $250 + $250 = $500 per month. Fable 5 on the identical workload costs $1,000. Sonnet 5 costs $300, and GPT-5.3 Codex costs $227.50.
That $500 figure is the crux of the whole review. You are paying $500 less than Fable 5 to give up 6.4 SWE-bench points, or paying $200 more than Sonnet 5 to gain 3.4. In our experience, the second trade is the one that matters: those 3.4 points concentrate in exactly the tasks that hurt — cross-file refactors, subtle state bugs, work where a failed attempt costs you an afternoon. The Fable premium, by contrast, only pays off at the genuinely hard tail. We run the full budget comparison in our cheapest coding models guide.
Its best seat on the team
Opus 4.8 is the rare model that can credibly sit in two chairs. As a builder, it is the most trustworthy implementer we have used short of Fable itself — hand it a well-scoped task and the diff comes back correct more often than anything at or below its price. As a planner or reviewer, it is entirely competent, though if budget allows we would rather put Fable 5 in the planning seat and let Opus execute.
The configuration we keep coming back to for teams who want one model everywhere: Opus 4.8 as the sole engine, planning and building alike. It is the strongest single-model setup that does not require flagship pricing. For teams running proper multi-agent stacks, Opus is the heavy builder you route the hard implementation tickets to while Sonnet 5 or Codex handles volume — the full role breakdown is in best model for each agent role.
Verdict
Opus 4.8 is the model we recommend most often, and the one we route the majority of our own real implementation work through. It lost the spotlight to Fable 5 by eleven days, but the value curve bends in its favor: it sits comfortably near the top of our best coding model rankings while costing like the upper-middle class, not the aristocracy.
Who should pay for it: professional developers, teams shipping production code, and anyone who wants one dependable model without maintaining a routing strategy.
Who should not: budget-first builders — Sonnet 5 or GPT-5.3 Codex will cover routine work at 40 to 55 percent of the cost. Terminal-heavy agent operators should look at GPT-5.5 first; the 78.9 vs 83.4 Terminal-Bench gap is real, and our cross-lab shootout covers exactly when it bites. And if you are wrestling problems at the bleeding edge, just pay for Fable. For everyone in between — which is most people — this is the one.