Claude Sonnet 5.5's Two-Week Usage Discount: What the 50% Actually Covers

The discount the vendor announced on 1 October 2026 does not appear on any invoice. It is not a price cut, not a promotional rate on tokens, and not something you opt into. From 11:00 AM PT on 1 October through 11:59 PM PT on 15 October, creating or editing an artifact in Claude or in cloud Claude Cowork makes a slice of the work that follows count at half rate against your five-hour session limit — and the model the vendor is pointing at while it runs is Claude Sonnet 5.5. The offer is documented, dated and automatic. What it does not do is nearly as specific as what it does, and the gap between the two is where most of the practical value sits.

The vendor's own framing of the mechanic is a single sentence: for a limited time, when you create or edit an artifact in Claude or Claude Cowork, a portion of the work that follows uses 50% less of your five-hour session limit. “A portion” is doing a lot of work there. The terms below are the vendor's, taken from the promotion's own help-centre page and not yet independently measured by anyone outside Anthropic — which is worth saying plainly, because a usage discount is unusually hard to verify from the outside. You cannot audit it from a bill. You can only watch your own session meter.

The two clocks that actually bind

The discount is a message count in chats and a step count in Cowork, and both are narrower than the phrase “the work that follows” suggests.

• In chats — the next 10 messages after Claude creates or edits an artifact in that chat are discounted. The 11th message is not, and nothing restarts the count except another artifact create or edit.

• Within a single reply — up to the first 15 steps are discounted. A step is a search, a tool call, or an artifact update. A long agentic reply that runs past 15 steps bills its remainder at the standard rate even though it is still inside the 10-message window.

• In cloud Cowork tasks — about the first 45 minutes are discounted if you start the task by choosing Document, Presentation or Design from the Output menu; otherwise, up to the next 80 steps after an artifact create or edit.

The interaction between the first two rows is the part that surprises people. Ten messages sounds generous until you notice that a single message can burn the whole step budget: ask for a substantial revision and the reply's sixteenth step onwards is full price, inside a window you reasonably believed was discounted.

How you open the chat decides whether your first message is discounted at all. Starting a new chat by choosing Document, Presentation or Design from the Output menu starts the discount on message one. Asking mid-conversation does not — the message where the artifact first gets created counts at standard rate, and the discount begins with the next one. If you know you are about to iterate on a deck, the difference between those two routes is one message and, in practice, the whole point of structuring it that way.

Who is in, who is out, and the line that catches people

The promotion is automatic for Pro, Max and Team plans. There is no code to enter and no setting to flip. It is not available on Free plans, and it is not available on Enterprise plans of either kind — Anthropic names both seat-based and usage-based Enterprise explicitly, which closes the obvious loophole of assuming a usage-based seat behaves like a Team seat.

The exclusions list is longer than the eligibility list, and a few entries matter more than the rest:

• Claude Code is excluded. A developer on a Max plan who lives in Claude Code gets nothing from this.

• The Claude API is excluded. This is a subscription-side promotion, not a model-price promotion.

• Claude in Slack is excluded, as are scheduled and unattended tasks, and Cowork tasks that run locally in Claude Desktop rather than in the cloud.

• Usage credits are excluded. If you have credits switched on and a message draws on them, that message bills at the standard rate regardless of where it sits in the window.

• Artifacts made with the earlier version of artifacts are excluded, as are Docs, Sheets and Slides files made through the Drive connector.

• The standalone Claude Design app is excluded — including projects you publish as artifacts from that app. Designs made inside Claude are included, so long as you pick Design from the Output menu in a chat or a Cowork task. Two things that both look like “Claude Design” land on opposite sides of this line.

One structural limit sits above all of it: the discount reduces what counts toward the five-hour session limit and leaves the weekly limit alone. A heavy week does not get rescued by artifact work. And the offer itself carries the usual vendor escape hatch — no cash value, not transferable, and Anthropic reserves the right to change or end it early, which is the reason to treat 15 October as a ceiling rather than a promise.

Why this is a Sonnet 5.5 story and not just a billing story

Anthropic paired the announcement with a model recommendation: give it a try with Claude Sonnet 5.5, which the vendor says has a strong eye for design and builds slides that need minimal editing. That line is marketing copy from the company that sells both the model and the subscription, and no independent evaluation of Claude Sonnet 5.5 on slide-generation or visual-design quality has been published that we can point to. Treat the design claim as a vendor position.

The model's other numbers are concrete. Claude Sonnet 5.5 shipped on 28 September 2026, three days before this promotion started, at $2 per million input tokens and $10 per million output tokens with cache reads at $0.20 per million, a 1,000,000-token context window and up to 128,000 output tokens. It is the cheaper half of Anthropic's current Sonnet-and-Opus pairing, and the placement is not accidental: the promotion asks you to spend your discounted messages on the model with the lowest per-token cost in the family, so the discount compounds with a rate you would have picked anyway. The same is true of the pairing it invites you away from — Claude Opus 5.5 and Claude Fable 5.1 both sit on materially higher cards, and the promotion does nothing to change that arithmetic.

The honest way to hold those two facts together is to keep them apart. The promotion is a documented, dated, automatic change to how a Claude subscription meters usage. The design capability is a claim. The first is why the promotion is worth reading about this week; the second is why Anthropic would like you to spend the discounted window on this particular model rather than on a dearer one.

What the discount is actually worth, roughly

Half of the usage for the ten messages after an artifact edit is not the same as half your usage for the day, and the arithmetic is easier to feel than to state. Ten discounted messages equal five standard ones. That is the entire saving per artifact event, before the 15-step ceiling starts trimming it further. A designer iterating on a presentation in twenty-message bursts gets the discount on the first ten of each burst and nothing on the rest.

Where the mechanic pays is in the shape of the work the vendor built it for: short, artifact-anchored loops. Draft a one-pager, turn it into a deck, mock up a matching visual, revise. Each create-or-edit restarts the count, so a session built out of repeated artifact revisions keeps landing inside the window, while a session built out of long open-ended analysis — no artifacts, no restarts — never touches it at all.

That is also the boundary between this promotion and the way most teams actually buy inference. If the work is going through the Claude API rather than the Claude app, none of the above applies. API usage is billed per token, and per-token billing does not have a five-hour session limit to discount. A team that has moved its document and deck generation onto an API pipeline gets no benefit from this offer at all — not because it is excluded on a technicality, but because the thing being discounted is a subscription meter rather than a price.

If you are building rather than clicking

For API work, the promotion is a non-event, and the interesting number is the one that was already there. anthropic/claude-sonnet-5.5 is carried on our own catalogue at $2.00 per million input tokens and $10.00 per million output, with cache reads at $0.20 and cache writes at $2.50 — the provider's own list price, passed through with no markup added, which is what makes a vendor rate change land on our side the same day it lands on theirs. The catalogue lists a 1,000,000-token context window with up to 128,000 output tokens, and serves it on both the chat-completions endpoint at /v1/chat/completions and the Anthropic-compatible messages endpoint at /v1/messages, so a codebase already written against the Anthropic SDK shape does not have to be rewritten to route through us.

Two things about that route are worth knowing before you point production at it. The first is that anthropic/claude-sonnet-5.5 is new on the catalogue — released 28 September 2026, and flagged as new rather than as an established route — which is exactly the situation automatic failover exists for, so a bad provider day does not become a bad deploy. The second is the comparison that this whole promotion invites. A subscription discount and a per-token rate are two different ledgers: the app's five-hour session limit can halve for two weeks while the API price does not move a cent, and a team that has outgrown the app should be comparing the pass-through rate to the cost of the seats rather than to the promotion. For workloads where the artifact is the deliverable and a human is in the loop, the subscription discount is the better deal for the next two weeks. For anything running unattended, it is irrelevant by design.

The one thing that is genuinely worth doing this fortnight is cheap: if you are on Pro, Max or Team and you have been drafting documents or decks in a plain chat, start the next one from the Output menu instead. That single choice moves your first message inside the discount, and it costs nothing but the click.

The catalogue lists a 1,000,000-token context window with up to 128,000 output tokens, and serves it on both the chat-completions endpoint at /v1/chat/completions and the Anthropic-compatible messages endpoint at /v1/messages, so a codebase already written against the Anthropic SDK shape does not have to be rewritten to route through us.