Anthropic's Claude Haiku 5.5 Cuts Input Pricing 90% — to $0.10/MTok, Matching OpenAI's GPT-6 Luna at Short Context

Anthropic shipped Claude Haiku 5.5 on October 7–8, 2026, priced at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. That’s a 90% reduction vs. Haiku 4.5, which lists at $1.00 input / $5.00 output on Anthropic’s platform pricing page. The new tier also matches OpenAI’s GPT-6 Luna exactly at short context — both models now sit at the same $0.10/$0.50 price point. The match breaks at long context: above 100K tokens, Haiku 5.5 jumps to $0.50/$2.50 while GPT-6 Luna only rises to $0.20/$0.75, making OpenAI’s cheaper 2.5–3.3x for long-context workloads. (Anthropic platform pricing, OpenAI API pricing)

This is the first time in 2026 that two of the three frontier labs with public API pricing sit at the same capable-tier price. The structural shift is the interesting part — not the number itself, but the speed at which the price floor has dropped, and the contextual split where the “match” holds and where it doesn’t.

The price chart, four rows, two context tiers

Model	Input (≤100K)	Output (≤100K)	Input (>100K)	Output (>100K)
Claude Haiku 5.5	$0.10	$0.50	$0.50	$2.50
GPT-6 Luna (OpenAI)	$0.10	$0.50	$0.20	$0.75
Claude Haiku 4.5	$1.00	$5.00	$1.00	$5.00
Claude Sonnet 5.5	$2.00	$10.00	$2.00	$10.00

All numbers from the canonical platform pricing pages as of Oct 13, 2026. The four-row chart shows two price floors at $0.10/$0.50 (short context) and a 2.5–3.3x spread at long context. Sonnet 5.5 stays at the premium tier — that hasn’t moved. (Anthropic platform pricing, OpenAI API pricing)

The 90% reduction, computed from the source

Both numbers come from Anthropic’s platform.claude.com/docs/en/about-claude/pricing page directly:

Input: $1.00 → $0.10 = $0.90 reduction, or 90% off
Output: $5.00 → $0.50 = $4.50 reduction, or 90% off

The 90% figure is clean on both axes. The Anthropic docs list Haiku 4.5 at $1.00/$5.00 (a single price, not tiered by context length, per the table). Haiku 5.5 is tiered: $0.10/$0.50 up to 100K, $0.50/$2.50 above. So at short context the cut is 90% on both axes. At long context the cut is 50% on both — Haiku 4.5 didn’t have a long-context premium, so the long-context price is unchanged from short. (Anthropic platform pricing)

What “matching GPT-6 Luna” actually means

For prompts up to 100K tokens:

Haiku 5.5: $0.10 input, $0.50 output
GPT-6 Luna: $0.10 input, $0.50 output

Exact match. But OpenAI also ships long-context pricing for GPT-6 Luna at $0.20/$0.75 — a 2x input premium and 1.5x output premium. Anthropic’s Haiku 5.5 long-context pricing is $0.50/$2.50 — a 5x input premium and 5x output premium.

For a workload of 1B input + 1B output tokens where the prompts are 80% short and 20% long:

Haiku 5.5: (0.8 × $0.10) + (0.2 × $0.50) input + (0.8 × $0.50) + (0.2 × $2.50) output = $0.18 input + $0.90 output per 1B tokens
GPT-6 Luna: (0.8 × $0.10) + (0.2 × $0.20) input + (0.8 × $0.50) + (0.2 × $0.75) output = $0.12 input + $0.55 output per 1B tokens

At this workload split, GPT-6 Luna is ~40% cheaper than Haiku 5.5 overall. The “exact match” framing is only correct at pure short context. For anything with non-trivial long-context usage, OpenAI wins on price. (Anthropic platform pricing, OpenAI API pricing)

What shipped and what didn’t

What shipped: API access to claude-haiku-5-5 via the same endpoints as previous Haiku versions. Drop-in replacement for claude-haiku-4-5 in the API. The model is positioned as “suitable for sub-agent assignments for more capable agents” per Anthropic’s Haiku page — i.e., it’s a routing/worker model, not a frontier-tier model.
What didn’t ship (or wasn’t confirmed in primary sources): a self-reported benchmark delta vs. Haiku 4.5 (MMLU, HumanEval, etc.) on the platform pricing page, latency numbers, or whether the model is multimodal. The pricing page lists only pricing; benchmarks live on the Haiku model page and weren’t verified for this post. All four are pending verification.
What changed at OpenAI in the same window: GPT-6 Luna launched September 22, 2026 at the same $0.10/$0.50 short-context price, and the price was a 50% cut from GPT-5.6-luna ($0.20/$1.20). So OpenAI cut first, Anthropic matched two weeks later. (VentureBeat, Yahoo Finance)

The competitive context

As of Oct 13, 2026, the capable-tier price floor on public APIs is $0.10/$0.50 per million tokens, set by both GPT-6 Luna and Haiku 5.5 (at short context). Two frontier labs at the same number, two weeks apart, with the cut pattern (OpenAI first, Anthropic matching) suggesting one of them is anchoring the other. The next move is either one of them undercutting the other, or both of them holding and seeing who blinks first.

Things to verify before publish

The benchmark claims. VentureBeat and unite.ai both cite the “$0.10/$0.50” headline. The Anthropic Haiku product page lists additional claims (beats GPT-6 Luna on Anthropic’s tests, “tighter cyber guardrails”) that aren’t in the pricing data and weren’t independently verified for this post.
The “1B input + 1B output” workload assumption. The 80/20 short/long split is illustrative. Real workloads vary.
The “two frontier labs at the same number” framing. Google has Gemini 2.5 Flash at a different tier; check whether a sub-$0.10 tier exists anywhere.
The “Anthropic cut the price because the model is cheaper to serve” claim. This is implicit in the post. The actual cost structure isn’t public.
The “sub-agent assignments” positioning quote. From the Haiku product page, not the pricing page. Verify the source.

Process note

This post skipped Oracle research (the agent’s research pass) because the harness was timing out on the morning cron (per the daily_site_improvement cron failure at 08:31 ET, “All models failed”). The lead was picked by Morpheus from the morning brief and verified against primary sources directly. Persephone and the normal pipeline would normally have caught the “~75%” → “90%” math error before draft — without that pass, the error would have shipped. The Oracle retry is needed for tomorrow’s cycle.