Claude 5.5 Pricing Hides Anthropic's Real AI Cost War

Anthropic's Claude 5.5 pricing refresh looks, on the surface, like a routine speed bump. The real story is what the company did not change: Sonnet 5.5 shipped at the exact same per-token price as its predecessor while Anthropic claims it now does nearly Opus-level work, and Opus 5.5 got a sticker-price cut that is smaller than the "cheaper to run" number the company is actually promoting. That gap between list price and real-world cost is where the next phase of the AI price war is actually happening.

The Headline Is Speed, the News Is What Anthropic Left Unchanged

Anthropic rolled out two new models in six days: Claude Opus 5.5 on September 22, 2026, and Claude Sonnet 5.5 on September 28, 2026, according to coverage aggregated from the launches (DeepLearning.AI and multiple outlets tracking the Sonnet follow-up). Anthropic's own current pricing page confirms both models' rates today, alongside Opus 5, Sonnet 5, and the rest of the Claude 4.x lineup still on sale.

Opus 5.5 is priced at $4 per million input tokens and $20 per million output tokens, a straightforward 20% cut from Opus 5's $5 input and $25 output. Sonnet 5.5, by contrast, kept Sonnet 5's exact $2 input / $10 output rate. Nothing moved on the list price for the model most developers actually use day to day.

Claude 5.5 Pricing Didn't Move for the Model Most People Use

That is the part that matters more than either launch date. Sonnet, not Opus, is the default tier for most API traffic because it is cheap enough to run on every request without a cost review. Anthropic holding Sonnet 5.5's Claude 5.5 pricing flat while claiming the model is "more than 30% faster" and "up to 30% less expensive for typical work" than Sonnet 5 is a bet that most customers don't need a discount, they need the same bill to buy more capability.

That bet lines up with how Anthropic frames the model's position: Sonnet 5.5 as the everyday workhorse for bug fixes, document drafting, and agentic coding, with Opus 5.5 reserved for the harder, more open-ended jobs. Multiple outlets reporting on the launch (inside.com.tw, itdaily.com, and letsdatascience.com among them) describe the same framing and the same unchanged $2/$10 rate.

The "40% Cheaper" Number Is Not the Price Cut You Think It Is

Here is the part most coverage glossed over. Anthropic's headline claim for Opus 5.5 was not the 20% list-price cut visible on the pricing page. It was a separate figure, repeated across several outlets, that Opus 5.5 is roughly 40% cheaper to run than Opus 5. That number is not a price-per-token comparison at all. It compares Opus 5.5 running at its default "medium" effort setting against Opus 5 running at "high," the setting most people actually left it on.

In other words, part of the savings comes from the model doing less work per query by default, not from the per-token price falling 40%. A team that pins Opus 5.5 to high effort to match its old Opus 5 behavior will see a 20% price cut, not a 40% cost cut. That distinction barely surfaced in the wave of launch coverage, most of which repeated the 40% figure without the asterisk.

Effort Dials, Not Price Tags, Are Where the Real Savings Live

This is the structural shift worth naming. Frontier labs used to compete almost entirely on sticker price and raw benchmark scores. Now the lever that actually moves a customer's bill is a per-request effort or reasoning-depth setting that most developers never touch after their first integration. Anthropic, OpenAI, and Google have all shipped some version of this dial over the past year, and it means two companies can run "the same model" at wildly different real costs depending on a setting buried in their API call.

That matters because it shifts the burden of cost control from the vendor to the buyer. A lab can advertise aggressive savings while leaving the actual per-token price nearly untouched, and whether a customer captures that savings depends on whether their engineering team goes back and re-tunes a default. Our earlier look at GitLab's AI credit caps found a similar pattern: the posted price is rarely the number that determines what a team actually pays.

Who Should Switch Tiers, and Who Should Sit Tight

For teams already paying for Opus on routine agentic coding or document tasks, it is worth testing Sonnet 5.5 first. Anthropic's own benchmark table shows Sonnet 5.5 trailing Opus 5.5 on most reported evaluations but rarely by a wide margin, and leading on at least one (Terminal-Bench 4.0, where Anthropic reports Sonnet 5.5 at roughly 70% versus Sonnet 5's roughly 10%). At half of Opus 5.5's per-token price, that is a reasonable default to try before paying for the larger model.

• Switch to Sonnet 5.5 if your workload is routine agentic coding, bug fixes, or document generation and you were defaulting to Opus mainly out of caution.
• Stay on Opus 5.5 for sustained, open-ended reasoning work where Anthropic and third-party evaluators still show a real quality gap.
• Re-check your effort or reasoning-depth setting before comparing bills across model generations; the default changed, and that changes your cost even if the price per token did not.
• Skip this entirely if you're a consumer on a flat Claude subscription; per-token pricing does not affect your bill.

This matters most to developers and engineering leads managing API spend across model tiers, and to anyone evaluating whether to renew an enterprise AI contract this quarter. Casual chatbot users on a flat monthly plan can safely ignore all of it; the per-token economics described here only bite on usage-based billing.

Self-Reported Benchmarks Still Need an Independent Check

The strongest pushback on any of this is simple: these are the vendor's own numbers. Anthropic reports Sonnet 5.5 scoring about 70.6% on Terminal-Bench 4.0. Independent evaluator Artificial Analysis, running its own version of the same test, reportedly measured Sonnet 5.5 at 63.6%, a meaningfully lower score on the same benchmark name. Neither number is necessarily wrong; benchmark implementations vary, and a seven-point gap is well within the range you'd expect from different harnesses and prompt formats.

The gap is a reminder that "Terminal-Bench 4.0" on a vendor's launch page and "Terminal-Bench 4.0" run by a third party are not guaranteed to mean the same test conditions.

That is a legitimate limitation, and it is the honest answer to anyone tempted to treat launch-day benchmark charts as settled fact. It does not undo the pricing comparison, which comes from Anthropic's own public pricing page and is independent of any benchmark dispute, but it is a reason to treat every performance claim in this piece, and in every other AI launch story, as a company's self-report until an independent lab replicates it.

What to Watch Next

Anthropic has already said Claude Haiku 5.5 would follow "in the coming weeks," and it is already listed on the company's pricing page at $0.10 input and $0.50 output per million tokens, completing a full repricing of the lineup in roughly a week. Watch whether OpenAI and Google respond with their own effort-dial tricks rather than straightforward price cuts; if the pattern holds across all three labs, "price cut" announcements in AI will keep getting harder to compare at face value. Our earlier coverage of Anthropic's IPO prospectus covers the margin pressure driving this behavior from the other side of the balance sheet.

The concrete takeaway: before you compare any two AI model generations on cost, check three things, not one. Check the list price per million tokens, check what effort or reasoning setting each number assumes, and check whether the performance claim behind it has been replicated by anyone other than the company selling the model.

Sources

• Anthropic/Claude official pricing page
• DeepLearning.AI, "Claude Opus 5.5 Leaps Forward"
• Inside.com.tw on the Claude Sonnet 5.5 launch
• ITdaily, "Anthropic introduces Claude Sonnet 5.5"
• Apidog benchmark roundup for Claude Sonnet 5.5