Anthropic Hid the Hike in the Tokenizer

Anthropic's new tokenizer is a stealth price hike: the per-million rate card on Sonnet 5 and Opus 4.8 didn't move, but identical code now bills as up to 73% more tokens than GPT-5.6 — and Playcode's July 14 audit found Opus 4.8's effective input cost behaves like $7.50 per million, not the published $5.00, weeks before Sonnet 5's introductory discount expires on August 31.

What's new:

• The Register broke the story July 14 after Playcode published cross-vendor tokenizer counts using each provider's own billing endpoints
• Anthropic shipped the tokenizer with Sonnet 5 in late June and Opus 4.8 before that — same list prices, denser token cuts
• Ploy's production migration to GPT-5.6 Sol reported 27% lower per-build cost and 35% fewer input tokens versus Opus 4.8 on identical website builds

How the hidden surcharge works:

• AI bills multiply two numbers: tokens your content becomes × price per token
• Pricing pages show only the second number and treat the first as fixed. It isn't
• Anthropic's docs acknowledge the new tokenizer produces "approximately 30% more tokens" for the same text on Sonnet 5 versus Sonnet 4.6
• The company called the tradeoff explicit: an "updated tokenizer that changes how the model processes text to improve performance" — with the same input mapping to "roughly 1.0–1.35×" more tokens depending on content type

The numbers that matter:

• On a 2,888-character TypeScript file, Playcode counted 1,178 tokens on Claude's new tokenizer versus 898 on Anthropic's old one (+31%) and 681 on GPT-5.6's o200k tokenizer (1.73× gap)
• Rust: 1.58×. JavaScript: 1.52×. Python: 1.50×. English prose: 1.40×
• Opus 4.6 and Opus 4.8 share a $5/$25 list price — but Playcode's blended coding workload makes 4.8's effective rate ~$7.50/$37.50
• Sonnet 5 after August 31: list $3/$15, effective ~$4.50/$22.50 — about one-third more than Sonnet 4.6 at the same sticker price

The intro discount is a bridge, not a cut:

• Sonnet 5 launched at $2/$10 through August 31, 2026 — which roughly offsets the extra tokens for now
• From September 1, standard $3/$15 pricing returns while the tokenizer stays inflated
• Anthropic's own migration guide tells developers to "recount prompts with token counting" and revisit max_tokens limits — admission that drop-in upgrades are not cost-neutral
• Adaptive thinking is now on by default for Sonnet 5, stacking thinking tokens on top of the denser input counts

Production proof from Ploy:

• Ploy ran hundreds of real website builds head-to-head before switching its default from Opus 4.8 to GPT-5.6 Sol
• Mean completed build: $3.06 on Opus versus $2.22 on GPT-5.6 (−27%), 8 minutes versus 3 minutes 42 seconds (2.2× faster)
• Input tokens: 2.60M versus 1.70M. Output tokens: 33.0K versus 17.1K
• Ploy's July 9 migration post: GPT-5.6 "finished pages 2.2× faster, cost 27 percent less, and used about half the output tokens"
• Playcode, which defaulted Pro users to Opus 4.8, warned the new tokenizer packs "roughly 25-48% more tokens into a typical request" at the same per-token price

Why enterprises should panic quietly:

• Token budgets are negotiated on rate cards. Tokenizers are buried in release notes
• Cache reads and writes bill per token too — a 32% tokenizer inflation taxes every cached agent session
• Uber capped AI spend at $1,500 per employee after burning its annual budget in four months. Palantir's Alex Karp called token billing "completely wrong" on CNBC July 1
• Anthropic is racing toward a public listing analysts peg near $1 trillion — revenue optics favor flat list prices over honest unit economics

What to do now:

• Run your own fixtures through each vendor's count_tokens endpoint before renewing contracts
• Treat any tokenizer change as a price change even when the rate card doesn't
• Benchmark dollars per completed task, not dollars per million tokens — Ploy's migration only looked expensive until cache keys were fixed

Convina's view: Anthropic did not raise prices. It redefined what a token is — then offered a two-month discount to make the swap feel free. That is pricing craft, not transparency, and it lands the same week 200 economists signed a letter admitting they cannot model AI's economic impact while OpenAI's CFO put her name on the warning. If the people selling the models cannot forecast the labor shock, the least they can do is quote an honest unit price. Until vendors publish effective cost per byte — or per completed task — enterprise buyers are comparing sticker rates while the scissors change underneath. The tokenizer is the hike. Read the release notes like a tariff schedule.