Sonnet 5, Opus 4.8, or Fable 5? Choosing Your Claude Code Model by Cost

Two things changed in Claude Code the same week, and together they made a question that used to be easy kind of hard:

• Claude Sonnet 5 launched (June 30) and is now the default model in Claude Code (v2.1.197) and on Free/Pro. Most of your fresh sessions are already running on it.
• Claude Fable 5 is back (restored July 1 after its export-control suspension) — with a new billing mechanic and a clock on it.

So now you’ve got three serious options — Opus 4.8, Sonnet 5, and Fable 5 — with different prices, different billing rules, and a tokenizer wrinkle that makes the sticker price misleading. Here’s the honest, cost-first way to think about it. (We’re a cost tracker, so this is about the dollars — not a claim about which model is “smarter.”)

Audit your team's Claude spend, free →

Sonnet 5: the new default, and cheaper per token — but read the asterisk

Sonnet 5 is $2/$10 per million tokens right now (input/output), stepping up to $3/$15 on August 31. That’s below Opus 4.8, and since it’s the new default, it’s already what most of your sessions use.

Here’s the asterisk that matters: a lower per-token price is not the same as a lower bill. Sonnet 5’s tokenizer tends to emit more tokens for the same work — often around 30% more. So the same task can use more tokens on Sonnet 5 than the raw “$2 vs Opus” comparison implies, and the real per-task saving is smaller than the rate card suggests. It’s still very likely cheaper for most work — just not by as much as the headline number, and not guaranteed on every task.

The good news for tracking: this tokenizer difference doesn’t dent your cost number at all. WTClaude prices Sonnet 5 off the model’s own reported token counts (and, in terminal Code, off Anthropic’s own computed cost), so more tokens simply show up as more tokens — no re-calibration, no drift. (Why a cost tracker can still disagree with your bill.)

Fable 5 is back — and there’s a ~July 12 clock

Fable 5 returned July 1. The billing works like this:

• It’s included for up to 50% of your weekly usage limit through ~July 12 (on Pro, Max, Team, and premium Enterprise plans). (That cliff was originally July 7; Anthropic extended it to ~the 12th after backlash — treat the date as movable and re-check the current one.)
• After that allowance — or on standard Enterprise, which gets no allowance — Fable bills as usage credits at $10/$50 per million (cache-read $1).
• Requests that get blocked reroute to Opus 4.8.

Two honest notes. First, “included” is an allowance cap, not free-unlimited — it’s up to half your weekly limit, then the meter starts. Second, any “what Fable will cost me after the cliff” number is a labeled estimate: it depends on your usage, and it assumes the announced $10/$50 holds. We can’t see how much of your weekly limit you’re burning, so we don’t pretend to know exactly how much the 50% allowance covers — we show what your Fable usage would cost as credits, clearly labeled.

(If you’re keeping a mental map of Claude’s billing changes, this Fable allowance cliff is its own thing — separate from the Agent-SDK pool split that was announced earlier and then paused. Different wallets; don’t let anyone blur them. Here’s the plain billing explainer.)

The choice is now genuinely hard — so price it against your usage

Opus 4.8 for the heavy reasoning, Sonnet 5 as the cheaper default, Fable 5 for the work it’s best at (on a clock, at 2× Opus rates once metered). The rate cards alone won’t tell you what you’d actually pay, because it depends entirely on how you work — which model you lean on, how much cache you hit, how token-heavy your sessions run.

That’s exactly the question WTClaude’s compare-models answers. It takes your own recorded, billing-grade usage and re-prices it across all three models, so you see what this month would have cost on each:

npx wtclaude setup
wtclaude compare-models

A few things we’re careful about, because they’re the difference between an honest tool and a misleading one:

• It re-prices your recorded usage — it is not “the same task run on each model.” A different model would emit different token counts for the same work (that tokenizer point again), so the comparison is “your real usage, re-priced,” clearly labeled as an estimate — not a lab benchmark.
• It’s split by where you work. The Code (terminal) row is billing-grade — Anthropic’s own computed cost, re-priced. The Cowork row is a labeled estimate. Chat is excluded (there’s no local cost data for it, so we say so rather than show a fake zero). Model choice is made per surface anyway, so this is the split that actually helps you decide.
• Cost, not quality. We surface what each option costs. Whether Sonnet 5 is “good enough” for your work versus Opus is your call — we’re not going to dress a cost tool up as a quality benchmark.

You also get this everywhere you already look: the same three-model comparison renders in the CLI, in the web dashboard tile, and in the desktop companion — driven off one shared calculation so the numbers can’t disagree. (Want to see how WTClaude stacks up against the other trackers first? Here’s the honest side-by-side.)

The one number you can trust underneath all of it

Model prices change, allowances expire, tokenizers differ — but the anchor doesn’t move: WTClaude reads the same cost number Anthropic’s own statusline computes for terminal Code. It’s Anthropic’s own figure, not a third-party reconstruction from broken logs (and not a guarantee of your final invoice — it’s Anthropic-computed, terminal-Code). Everything else — the Fable forecast, the cross-model comparison, desktop and Cowork estimates — is clearly labeled as an estimate, so you always know which numbers are solid and which are projections.

That’s the whole idea: when the model landscape gets confusing, the tracker’s job is to make your costs legible, honestly. If you want the full picture of how Claude Code billing fits together — pools, credits, and limits — start here.

FAQ

Is Claude Sonnet 5 cheaper than Opus 4.8? Per token, yes — $2/$10 (intro, → $3/$15 on Aug 31) undercuts Opus. But it’s not a guaranteed saving: Sonnet 5’s tokenizer emits more tokens for the same work (often ~30% more), so the raw per-million gap overstates the real saving. Price it against your own usage with wtclaude compare-models.

How much does Claude Fable 5 cost after July 12? Included up to 50% of your weekly limit through ~July 12, extended from the original July 7 date (Pro/Max/Team/premium Enterprise), then usage credits at $10/$50 per million (cache-read $1); standard Enterprise is credits-only. Any post-cliff figure is a labeled estimate that depends on your usage and assumes $10/$50 holds.

How do I compare Claude models by cost? wtclaude compare-models re-prices your recorded usage across Opus 4.8, Sonnet 5, and Fable 5 — split by surface (Code billing-grade, Cowork a labeled estimate, Chat excluded), every projection labeled. It compares cost, not quality.

Does the Sonnet 5 tokenizer make my cost tracking wrong? No. WTClaude reads the model’s own reported token counts (and Anthropic’s computed cost in terminal Code), so more tokens just register as more tokens — no drift. The tokenizer only matters as a caveat on cross-model projections.

Is WTClaude affiliated with Anthropic? No — it’s an independent, free, open-source project, not affiliated with Anthropic.

---

WTClaude is a free, open-source, billing-grade Claude Code cost tracker. See what your month would cost on each model: npx wtclaude setup, then wtclaude compare-models. Numbers are billing-grade for terminal Code (Anthropic-computed) and clearly-labeled estimates everywhere else.