Anthropic launched Claude Sonnet 5 on June 30, 2026, pricing it at $2 per million input tokens and $10 per million output tokens — undercutting Sonnet 4.6's $3/$15 rate by roughly a third. If you're an engineer or technical PM budgeting API spend right now, here's the math before you flip the model string in production. Below the September 1, 2026 pricing cutover, Sonnet 5 wins on cost at every workload bucket. Above it, Sonnet 4.6 wins by a margin that scales with your volume — because Sonnet 5's nominal price reverts to Sonnet 4.6's exact rate, but its tokenizer burns roughly 30% more tokens for the same text.

TL;DR: the verdict

Workload	Sonnet 4.6 cost/mo	Sonnet 5 cost/mo (now-Aug 31)	Sonnet 5 cost/mo (from Sep 1)	Winner
Light (100 prompts/day, ~50K in / 10K out tokens daily)	$6.60	$5.72	$8.36	Sonnet 5 now, Sonnet 4.6 after Sep 1
Medium (1,000 prompts/day, ~500K in / 100K out tokens daily)	$66.00	$57.20	$85.80	Sonnet 5 now, Sonnet 4.6 after Sep 1
Heavy (10,000 prompts/day, ~5M in / 1M out tokens daily)	$660.00	$572.00	$858.00	Sonnet 5 now, Sonnet 4.6 after Sep 1

Short answer: switch to Sonnet 5 now and you save 13-15% through August 31, 2026. Stay on it past that date without re-checking the math, and you pay 26-30% more than Sonnet 4.6 for the same workload.

What each one actually costs

Claude Sonnet 4.6 pricing breakdown

Standard API: $3/MTok input, $15/MTok output — usage-based, no subscription tier, per Anthropic's pricing page.
Prompt caching: 5-minute cache writes at $3.75/MTok, 1-hour writes at $6/MTok, cache hits at $0.30/MTok.
Batch API: $1.50/MTok input, $7.50/MTok output — a flat 50% discount for async jobs.

No hidden costs here: no seat minimums, no annual lock-in, no overage fees. You pay for the tokens you send and receive, full stop.

Claude Sonnet 5 pricing breakdown

Intro API (through Aug 31, 2026): $2/MTok input, $10/MTok output, confirmed on Anthropic's Sonnet 5 launch post and the same pricing docs.
Standard API (from Sep 1, 2026): $3/MTok input, $15/MTok output — identical nominal rate to Sonnet 4.6.
Batch API: $1/MTok input, $5/MTok output during the intro window; $1.50/$7.50 after.

The hidden cost: Sonnet 5 runs on a new tokenizer that, per Anthropic's own pricing documentation, "produces approximately 30% more tokens for the same text" (range cited as 1.0x-1.35x depending on content type). TechCrunch's launch coverage frames Sonnet 5 as "a cheaper way to run agents," and that's true through August — but the per-token discount was set deliberately to make the transition "roughly cost-neutral" against the token-count increase, not to deliver a permanent discount. This isn't an isolated case — a June 2026 analysis of OpenAI, Anthropic, and Google pricing counted 14 combined pricing changes across the three vendors between January and June 2026, several of them undisclosed rate increases hiding behind a model upgrade.

Break-even, walked through

Take the Medium bucket: 1,000 prompts/day, ~500K input tokens and 100K output tokens daily, 22 working days a month. On Sonnet 4.6, that's (500,000 × $3 + 100,000 × $15) / 1,000,000 × 22 = $66.00/mo. On Sonnet 5 during the intro window, the same text now tokenizes to roughly 650,000 input and 130,000 output tokens (the ~1.3x multiplier), priced at $2/$10: (650,000 × $2 + 130,000 × $10) / 1,000,000 × 22 = $57.20/mo — a $8.80/mo saving, about 13%.

Run that same calculation after September 1, 2026, when Sonnet 5's price reverts to $3/$15: (650,000 × $3 + 130,000 × $15) / 1,000,000 × 22 = $85.80/mo — $19.80/mo more than staying on Sonnet 4.6, a 30% increase. The inflection point isn't a workload threshold here; it's a date. Every bucket flips the same direction on the same day, because the tokenizer multiplier scales with volume while the price gap between $2 and $3 doesn't survive the cutover.

What switching actually costs in time

Migration time: 2-4 hours — update the model identifier in your API calls or SDK config, then re-run your eval suite against production prompts to confirm output quality holds at the new model.
Ramp period: 0-1 day. This is an API model swap, not a tool swap — no UI to relearn, no team retraining, just a redeploy and a monitoring window.
Lock-in to leave: none. Both models bill on pay-as-you-go usage with no subscription, no annual contract, and no seat minimums — switching back is the same one-line config change.
Recovery: at Medium workload, the engineering time pays for itself inside the first day of usage, since the saving is $8.80/mo against roughly 2-4 hours of one-time work. The real cost isn't the switch — it's forgetting to switch back, or re-check the math, before September 1.

Pick by your profile

Solo dev, side projects, <500 reqs/day: switch now. At Light workload the saving is only $0.88/mo, but there's no downside risk and no migration cost worth worrying about — just set a calendar reminder for August 31.
Team of 5-20, predictable workload: switch now, but put the September 1 cutover on your cost-monitoring dashboard. At Medium workload you're looking at a swing from -$8.80/mo to +$19.80/mo on the same line item.
Cost-sensitive batch processing: the batch discount stacks the same way — $1/$5 intro vs $1.50/$7.50 standard — so the same 30%-token-count math applies. Model the cutover before committing to Sonnet 5 for high-volume async jobs. If you're weighing non-Anthropic alternatives for batch workloads, see our breakdown of Claude Opus vs DeepSeek pricing.
Latency- or quality-critical user-facing workloads: the price delta at any bucket here is small change against engineering time. Pick on output quality and agentic accuracy first, and let the per-token rate be the tiebreaker, not the deciding factor. Our GLM-5.2 vs Sonnet 4.6 comparison walks through a similar tradeoff against a non-Anthropic model.

FAQ

Is Sonnet 5 actually cheaper than Sonnet 4.6?

Yes, through August 31, 2026 — by about 13-15% at typical workloads. From September 1, 2026, the nominal price reverts to Sonnet 4.6's rate while the new tokenizer still uses ~30% more tokens for the same text, making Sonnet 5 the more expensive model at every workload bucket from that date.

How long until switching pays for itself?

Almost immediately. Migration is a one-line model-identifier change plus 2-4 hours of eval re-runs — there's no dollar cost to recover, since both models bill the same pay-as-you-go way with no subscription fee or lock-in.

What if my workload changes?

The math scales linearly with volume, so the percentage gap holds regardless of bucket size: monthly cost = (daily input tokens × input price + daily output tokens × output price) / 1,000,000 × 22 working days, with Sonnet 5's token counts multiplied by roughly 1.3x to reflect its tokenizer. Re-run that formula with your own token counts before committing either way.

Are these prices current as of June 2026?

Pricing pulled from 3 sources published or last updated on June 30, 2026: Anthropic's official pricing documentation, Anthropic's Sonnet 5 launch announcement, and TechCrunch's coverage of the launch. Anthropic can change pricing without notice — check the official pricing page before committing budget.

#claude-sonnet-4-6

Claude Sonnet 4.6 to Sonnet 5: Should You Switch in 2026?

#claude-sonnet-4-6

Claude Sonnet 4.6 to Sonnet 5: Should You Switch in 2026?