
Gemini 4 Neon vs Claude Haiku 5.5: The Ends of One Ladder
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 184 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 727 tok/s
- AlibabaNEWQwen: Qwen4 Max2026-11-16$1.50 / $4.50 per 1M tokens · 407 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 150 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 99 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 83 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 121 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 108 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 67 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 90 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 41 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 212 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 111 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 160 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
Put Gemini 4 Neon and Claude Haiku 5.5 next to each other on a spec sheet and the comparison looks like a mismatch: thirty times the input price, thirty times the output price, and a model that costs a tenth of a cent per thousand tokens against one that costs three. Then you notice that both models are sold as ladders rather than as single points. Neon has four thinking levels and defaults to high; Haiku 5.5 has five and defaults to medium. Once you compare rung to rung instead of card to card, the thirty-fold ratio stops being the interesting number and the shape of the two ladders becomes the thing worth arguing about.
Both models are current. Claude Haiku 5.5 reached general availability on 7 October 2026 at $0.10 per million input tokens and $0.50 per million output; Gemini 4 Neon shipped on 18 November 2026 at $3.00 and $15.00. Both carry a 1M-token context window, both accept text and images, and both charge $0.01 and $0.15 respectively for cached input — the same ninety-five percent discount applied to very different base rates.
Two ladders, nine settings
Artificial Analysis scores every rung of both models on the same nine-part evaluation set — AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Omniscience v1.1, GDP.pdf, HLE-Diamond, CritPt v2 and Terminal-Bench-Science 0.1 — and publishes a per-task cost for each. That gives nine comparable points rather than two, and the per-task cost is measured rather than derived from the rate card, so it already folds in how many tokens each model chooses to spend.
• Neon at low — 44.1, $0.75 per task.
• Neon at medium — 54.9, $1.90.
• Neon at high (default) — 59.6, $3.74.
• Neon at xhigh — 61.4, $6.44.
• Haiku 5.5 at low — 18.6, $0.044.
• Haiku 5.5 at medium (default) — 24.6, $0.092.
• Haiku 5.5 at high — 29.9, $0.159.
• Haiku 5.5 at xhigh — 34.6, $0.248.
• Haiku 5.5 at max — 36.5, $0.411.
Two things fall out of the list immediately. Neon's cheapest rung out-scores Haiku's most expensive one by 7.6 points, and Neon's most expensive rung costs 15.7 times Haiku's most expensive one rather than the 30 times the rate cards imply. Neither of those is a rounding artefact; both come from the same place, which is that the two models spend tokens very differently at the top of their ranges.

What 'High' means on each card
The vocabulary is the first trap. On Neon, high is the rung the model ships with — the one you get if you pass no thinking parameter. On Haiku 5.5, high is the third of five and the default is medium. So a comparison written as "Neon high versus Haiku high" is comparing a default to a non-default, and it produces a 29.7-point gap that neither vendor would recognise as the headline.
Compare the actual defaults — Neon at high, Haiku at medium — and the gap widens to 35.0 points, because Haiku's default sits one rung lower than its name suggests. Compare the tops of the two ladders — xhigh against max — and it narrows to 24.9 points. That is a ten-point swing driven entirely by which pair of settings you decided to call the comparison, and it is wider than the gap between several adjacent models on the same leaderboard.
Matched rungs, and what the gap does
Line the rungs up by name and the index gap moves in a narrow band, from 25.5 to 30.3 points, with no trend in either direction:
• At low — Neon 44.1 against Haiku 18.6, a 25.5-point gap, 17.0x the cost per task.
• At medium — 54.9 against 24.6, 30.3 points, 20.8x.
• At high — 59.6 against 29.9, 29.7 points, 23.6x.
• At the top — 61.4 against 34.6, 26.8 points, 25.9x.
The intelligence gap is stable. The cost ratio is not: it climbs from 17x to 26x as you walk up the ladder, because Neon charges more per additional point of thinking than Haiku does. Pushing Haiku from high to xhigh buys 4.7 index points for 56 percent more per task. Pushing Neon from high to xhigh buys 1.8 points for 72 percent more. At the margin, the cheaper model is the one with the better return — which is the opposite of what a thirty-fold rate-card ratio would lead you to expect.
The fifth rung is where the claim breaks
Haiku 5.5 has a rung Neon does not: max. It is also the worst-value rung either model sells. From xhigh to max, Haiku gains 1.85 index points and its per-task cost rises from $0.248 to $0.411 — a 66 percent increase for a 5 percent improvement. In the same step its measured time to first answer token jumps from 46.2 seconds to 180.9, roughly four times.
That is not a defect unique to Haiku. It is what an extra reasoning tier looks like when the harness has run out of easy wins: the model keeps thinking and stops improving. But it does mean that anyone quoting Haiku 5.5's 36.5 is quoting the one configuration of that model that costs four times as much latency as the one below it, and anyone quoting Neon's 61.4 is quoting the same kind of ceiling for the same kind of reason. Both headline numbers are the top rung of a ladder that gets steep at the end.

Reasoning tokens are the hidden column
The per-task cost numbers make sense once you look at how much each model writes to itself. On the index runs, Neon at xhigh produces about 80,500 reasoning tokens and 61,100 answer tokens per task. Haiku 5.5 at max produces about 172,800 reasoning tokens and 53,200 answer tokens.
So the model with the thirty-fold rate card writes less than half as much reasoning as the model with the ten-cent one, at the top of both ladders. Haiku compensates for a smaller model by thinking longer; Neon compensates for a higher price by thinking less. That is why the per-task ratio of 15.7x sits so far below the 30x the price list suggests, and why the ratio is widest at high — the rung where Neon's token habit is closest to Haiku's and its rate card is doing all the work.
The other column in the same panel is fabrication. On AA-Omniscience, Neon's hallucination rate at its default is 13.9 percent; Haiku 5.5's at its default is 40.4 percent, and 34.2 percent at max. That is a two-to-three-fold difference in how often the model asserts something the source does not support, and unlike the index it barely improves as Haiku climbs its ladder. If your workload is the kind where a wrong fact is expensive to catch, the index gap understates the case for Neon rather than overstating it.
Where the cheap model stops being cheap
Haiku 5.5's price advantage is real and it is largest where the work is short. A classification call, a routing decision, a rewrite of a paragraph, an extraction from a form — anything where output is a few hundred tokens — costs fractions of a cent on Haiku and something closer to a cent on Neon. At that size the 30x ratio is close to literal and the index gap is irrelevant, because both models clear the bar easily.
The advantage narrows as the task lengthens, and it narrows for two reasons that compound. Neon's per-task cost is dominated by input, cache reads and cache writes — on the index runs, input accounting is 64 percent of the bill at high — so a long, stable prefix moves Neon's effective rate toward its $0.15 cache price. Haiku gets the same 95 percent cache discount, but on a base rate that is already tiny, so there is less to save. And on work that requires several steps of planning, Haiku's ladder tops out below where Neon's default sits, which means no amount of cheap tokens buys you the answer.
There is a third effect worth naming because it is easy to miss: the measured briefcase Elo — a different harness, scored separately from the index — actually favours Haiku at the top by a narrower margin than the index does. Neon xhigh scores 1699 with a 95 percent interval of 1688 to 1710; Haiku max scores 1572. At the default rungs the gap is 1632 against 1441. So a second, independent harness agrees on the direction and disagrees on the size, which is the normal state of affairs and a reason to check two leaderboards before standardising.
And there is a fourth, which is the one the rate card hides completely. Haiku 5.5 prices in two bands: $0.10 and $0.50 per million up to a 100,000-token prompt, then $0.50 and $2.50 above it, with the cache read moving from $0.01 to $0.05. The threshold counts the whole input including cache reads and writes, so a long context trips it easily. At that point Haiku's input price is five times what the headline says, its output price is five times, and the ratio to Neon's flat $3.00 and $15.00 — which steps at no threshold at all — is no longer thirty-fold but six-fold, or three-fold on output. Anyone comparing the two cards at face value is comparing a two-band price to a one-band price and treating the first band as the whole story.
A decision rule that survives the rungs
Four rules, in the order that matters:
• If the task is short and the answer is checkable, take Haiku 5.5 at medium or low and stop reading. The index gap does not bind at that size and the price gap is enormous.
• If the task needs multi-step planning or repo-level code changes, Haiku's ceiling is the problem, not its price. Its best rung scores 36.5; Neon's default scores 59.6. You are not trading cost for quality at that point, you are trading a working answer for a cheap one.
• If you are already on Neon, do not default to xhigh. The step from high to xhigh costs 72 percent more per task for 1.8 index points and 16 seconds of extra time to first token. Most pipelines are better served by medium, which is 54.9 for half the price of the default.
• If you are on Haiku, do not assume max. It is the fourth-best value on either ladder, and the gap between it and high is 6.6 index points for 2.6 times the cost.
Neither model is in OrcaRouter's catalogue today, so none of this is a pitch to move either one onto our key. What we do serve, at provider list price with no markup, is Claude Haiku 4.5 at $1.00 and $5.00 per million and Gemini 4 Carbon at $2.00 and $10.00 — the predecessors on each side. If the rung arithmetic above is the kind of thing you want to test on your own traffic rather than on a leaderboard, the useful move is to run the cheap rung of a model you can already call against the expensive rung of the one you are considering, on one key, with failover underneath in case the cheap rung turns out not to clear the bar.
The honest summary is that these two models are not really competing. One is a frontier-tier planner sold at a premium and the other is a high-volume worker sold near the floor, and the fact that their ladders overlap in name rather than in range is what makes the comparison feel closer than it is. Compare the ends — Neon's medium against Haiku's max — and you get 54.9 for $1.90 against 36.5 for $0.41. Eighteen points for four and a half times the money is a decision about what the task is worth, not a decision about which model is better.

