Claude Haiku 5.5 Is in GitHub Copilot Now: What the Cheapest Anthropic Model Costs Inside a Credit Budget

Claude Haiku 5.5 appeared in the GitHub Copilot model picker on October 7, 2026 — the same day the vendor released it — and the changelog entry that announced it spends most of its length on surfaces. Visual Studio Code, Visual Studio, the Copilot CLI, the GitHub Copilot cloud agent, the Copilot app, github.com, GitHub Mobile on iOS and Android, the JetBrains IDEs, Xcode, Eclipse. The sentence that matters if you hold the budget is shorter: the model is billed "at provider list pricing under usage-based billing." Inside Copilot that means AI credits, and Claude Haiku 5.5's rates are low enough that the credit arithmetic becomes the interesting part of the launch rather than an afterthought. This piece runs that arithmetic, separates what GitHub has actually demonstrated from what it has merely asserted, and covers two carve-outs in the vendor's own documentation that the launch coverage mostly skipped.

What GitHub shipped, and to whom

The entry is titled, plainly, "Claude Haiku 5.5 in GitHub Copilot," and it went up at 13:12 Pacific on October 7. It is available to Copilot Pro, Pro+, Max, Business and Enterprise subscribers — which is to say every paid tier, and notably not the free tier, where model selection is still handled automatically rather than by the user. The rollout is gradual, so the picker in your editor may lag the changelog by days; that is stated in the entry rather than left for you to discover.

GitHub's own framing of what the model is for is worth reading closely, because it is narrower than the vendor's:

• Fast, high-volume work — subagents, quick edits, and terminal tasks, in GitHub's own words • Not positioned as a replacement for the large models in the picker, but as the cheap tier underneath them • Generally available, per GitHub's supported-models reference, which lists Claude Haiku 5.5 as generally available, with a status of GA • Also listed on that reference with an IDE-version row marked TBD, which is the practical reason the rollout reads as gradual from the editor side • Priced at provider list pricing under usage-based billing, rather than at a Copilot-specific markup img src="2.png"> p>The surrounding changelog dates give the cadence context that a single model page cannot. Claude Sonnet 5.5 landed in Copilot on September 28, Claude Opus 5.5 on September 22, and Claude Fable 5.1 on September 1. Anthropic is shipping a Copilot integration roughly every two to four weeks and GitHub is turning each one around in a matter of days. That cadence is also why there is a clock on this: GitHub posted a deprecation notice in mid-September for "selected GitHub Copilot models" in mid-October, and then a follow-up on October 2 confirming the removals. If your team is pinning a model in a Copilot configuration, the pin is now the thing likeliest to break, not the price.

The rate card, read as a credit ledger

Claude Haiku 5.5 bills on two tiers, switched by request size rather than by any setting you choose. Everything below is GitHub's own published rate card, and all of it is per million tokens:

• At or under 100,000 tokens in — $0.10 input, $0.01 cached input, $0.125 cache write, $0.50 output • Above 100,000 tokens in — $0.50 input, $0.05 cached input, $0.625 cache write, $2.50 output • Claude Haiku 4.5, for contrast — $1.00 input, $0.10 cached input, $1.25 cache write, $5.00 output • Claude Sonnet 5.5, the tier above — $2.00 input, $0.10 cached input, $2.50 cache write, $10.00 output • Cache writes are billed separately on Anthropic models, above and beyond cached input — a line item that does not exist on every vendor in the picker img src="3.png"> p>Two things fall out of reading those rows together. First, Claude Haiku 5.5's short tier is a round ten times cheaper than Claude Haiku 4.5 on both input and output, and roughly twenty times cheaper on input than Claude Sonnet 5.5 — which is the whole reason it belongs in subagent slots, where a fleet of small calls multiplies whatever you pay per call. Second, the long-context tier is exactly five times the short one on every rate at once, at the same 100,000-token line. A 99,000-token request and a 101,000-token request are not 2 percent apart on the tokens that crossed the line; they are five times apart on all of them. That step is inherited straight from Anthropic's list price, and Copilot's credit system does not smooth it.

Now the credits. Copilot's usage-based billing converts model usage into AI credits at a documented rate of $0.01 per credit, and each plan carries a monthly pool:

• Copilot Pro, $10 a month — 1,500 credits, quoted as 1,000 base plus 500 flexible • Copilot Pro+, $39 a month — 7,000 credits, quoted as 3,900 base plus 3,100 flexible • Copilot Max, $100 a month — 20,000 credits, quoted as 10,000 base plus 10,000 flexible • Copilot Business, $19 per seat per month — 1,900 credits per user • Copilot Enterprise, $39 per seat per month — 3,900 credits per user • Organisation and enterprise pools are shared across seats rather than siloed per user • Usage beyond the pool is billed at $0.01 per credit — the same conversion, applied to overage • Code completions and next-edit suggestions sit outside credit billing entirely and remain unlimited on paid plans p>Put the two together and the shape of the launch becomes concrete. Take a moderately sized agent step — 40,000 input tokens, 2,000 output tokens, no caching. On Claude Haiku 5.5's short tier that is 40,000 times $0.10 per million plus 2,000 times $0.50 per million, or $0.005, which is half a credit at the documented conversion. The identical call against Claude Haiku 4.5's rates comes to $0.05, or five credits. The arithmetic here is ours, not GitHub's, but the inputs are all from the published rate card, and the conclusion is not subtle: a Pro subscriber's 1,500-credit pool absorbs roughly 3,000 such calls on the new Haiku where it absorbed roughly 300 on the old one. The pool gets ten times longer without the subscription changing.

That is the honest value proposition, and it comes with the honest caveat that the pool is quoted in credits rather than dollars, so a plan's real ceiling moves whenever a rate card moves. A tenfold cut on one model does not extend a pool tenfold if your traffic is spread across three.

What GitHub says the model does, and what it has not shown

The changelog contains one performance claim, and it is doing more work in the coverage than it can bear. GitHub writes that in early testing Claude Haiku 5.5 "matched Claude Sonnet 5 on many coding tasks" while using "significantly fewer tokens and steps."

Read the qualifiers in order. It is early testing. It is GitHub's testing, on GitHub's tasks, and GitHub is the party selling the integration. No harness is named, no task set is published, no figure accompanies either half of the sentence, and "many" is not a number. The claim may well be true — a small model matching a large one on narrow, tool-shaped coding tasks is a plausible result in 2026 — but as published it is a vendor assertion, and it should be labelled as one every time it is repeated. The "fewer tokens and steps" half is the more durable of the two, because it is a claim about efficiency rather than capability, and efficiency at the Haiku price point shows up in your own usage telemetry within a week. The capability half wants a third-party harness before it goes in a slide.

Anthropic's own numbers are a separate claim from GitHub's and carry the same provenance problem. Anthropic states that Claude Haiku 5.5 is about 75 percent cheaper to run on average, and that roughly 90 percent of requests previously sent to a Haiku-class model come in under 100,000 tokens, which is where the model is "especially good value." Both are the vendor describing its own product's economics. They are consistent with the published rate card — a large short-tier cut, a smaller long-tier one — and they are useful as vendor guidance rather than as a measured result.

Turning it on is an admin decision, and the default is yes

The part of the changelog that Copilot Business and Enterprise buyers should read twice is the access control paragraph. Administrators manage model access through the model policy in Copilot settings, and GitHub's documented default behaviour is that new models are enabled automatically unless an administrator has turned off the global default or explicitly disabled that specific model. The default is permissive; the restriction is opt-out.

Two consequences follow. If your organisation has strict model governance, Claude Haiku 5.5 may already be selectable by your developers without anyone having approved it by name. And if instead your organisation has already disabled the global default, the model will not appear for you no matter what the changelog says — which is the most likely explanation for the most common support question in the first week of any Copilot model launch. Before you debug a missing picker entry, check the model policy.

Two carve-outs Anthropic's documentation states that the launch posts did not

Both of these come from Anthropic's platform documentation rather than from any press coverage, and both affect what you can build rather than what you pay:

• Priority Tier is not available on Claude Haiku 5.5 — Anthropic's service-tier documentation names the model in its explicit exclusion list, alongside Claude Sonnet 5.5, Claude Opus 5.5, Claude Fable 5.1 and the Mythos line. If your architecture assumes provisioned throughput on the cheap tier, that assumption does not hold here. • The full 1 million token context window at standard pricing is granted to "Claude 4.6 and later models (except Claude Haiku 5.5)" — the parenthetical is Anthropic's, and Claude Haiku 5.5 is the one current model carved out of it. The window is still advertised on the model's spec row as 1M, with a 128,000 token maximum output; what the carve-out means is that you should not read the blanket long-context pricing guarantee as applying to this model. • The tokenizer changed. Anthropic states that Claude 4.7 and later models use a newer tokenizer producing approximately 30 percent more tokens for the same text, which means the per-token price and the per-page-of-your-text price moved in opposite directions. Claude Haiku 5.5's spec row lists "Fastest" as its position, an adaptive thinking default of medium, a June 2026 knowledge cutoff, and a model ID of claude-haiku-5-5.

None of these are dealbreakers for the subagent and quick-edit workloads GitHub is aiming the model at. They are the difference between quoting a launch post and quoting the documentation, and the second is what survives contact with a load test.

If you would rather route it than switch tools

There is a reasonable case for calling Claude Haiku 5.5 from your own stack rather than only through a chat-and-editor subscription: you get the raw short-tier rates without a credit conversion in between, and you keep the routing decision in your own code. The availability picture, stated plainly, is that Anthropic sells it through the Claude Platform and through Amazon Bedrock, Vertex AI and Microsoft Azure — and it is not on our catalogue yet. OrcaRouter carries Claude Haiku 4.5, Claude Sonnet 5.5, Claude Opus 5.5 and Claude Fable 5.1 today across a catalogue of just over 200 models, and Claude Haiku 5.5 is not among them. For the model itself right now, call Anthropic or one of the three clouds directly.

Where we do fit is the layer around it. OrcaRouter passes provider list price through at zero markup, so when a vendor repriced a model — as Anthropic did on this one — the new rate is live on our side the same day rather than after a billing cycle. Automatic failover means a cheap-tier call that errors lands on a working route instead of failing the request, which matters more for a subagent fleet making thousands of small calls than for a single interactive session. And the routing DSL exists precisely for the shape of decision this launch creates: send short, high-volume work to whichever model is cheapest below the threshold and long-context work to whichever is cheapest above it, from one endpoint on one bill, instead of encoding that split into two integrations. If Haiku 5.5 arrives on the catalogue, it slots into that policy without a code change.

What to do with this before the next sprint

The narrow reading is that Anthropic's cheapest current model is now selectable in every paid Copilot tier and is roughly ten times cheaper per token than the Haiku it replaces on the short tier. If your team is running subagents or terminal-shaped automation through Copilot, switching the model is a settings change with a large and immediate effect on how far a monthly credit pool stretches.

The two things worth checking first are the ones the announcement buries. Look at the 95th percentile of your request sizes: if it sits above 100,000 tokens, the tier you are actually billed on is the second one, and the tenfold comparison is a fivefold one. And look at whether your organisation's model policy has the global default switched on — for Business and Enterprise seats, that single setting decides whether this launch is available to your developers at all.

passes provider list price through at zero markup