GPT-6 Luna vs Claude Haiku 5.5 vs GLM-5.3 Flash: The Under-$1 Tier, Compared

Anthropic released Claude Haiku 5.5 on October 7 and priced it for volume: $0.10 per million input tokens and $0.50 per million output for prompts up to 100K tokens — roughly 90% less than Haiku 4.5. That puts it in the same price band as GPT-6 Luna ($0.20/$1.20) and GLM-5.3 Flash ($0.15/$0.50), and for the first time Anthropic published direct head-to-head benchmarks against Luna. This is a routing guide: what the published numbers show, how Haiku's 100K-token pricing cliff changes the math, and when each of the three is the right default.

Pricing and Specs

Claude Haiku 5.5: two-tier pricing. Prompts up to 100K tokens: $0.10 input / $0.50 output per million. Over 100K: $0.50 / $2.50 — a 5× jump. Cache reads $0.01 (up to 100K) or $0.05 (over); cache writes $0.125 / $0.625. 1M context window. Anthropic's fastest model to date — up to 234 tokens/second at max effort. Text and image input, adjustable effort setting from low to max — a first for the Haiku class. GPT-6 Luna: $0.20 input / $1.20 output per million tokens, flat across the whole context window. $0.02 cached input (10× discount). 1.05M context, 128K max output. Full OpenAI tool suite — hosted shell, computer use, MCP, web search, apply patch — plus reasoning-effort controls. GLM-5.3 Flash: $0.15 input / $0.50 output per million tokens, $0.03 cached input. 320B-parameter mixture-of-experts (18B active per token). 1M context. Native multimodal input — text, image, and video in the same call. MIT-licensed open weights.

The Head-to-Head Anthropic Published

Anthropic's Haiku 5.5 launch page benchmarked it directly against GPT-6 Luna. The published numbers: GDPval-AA v2.1 (real-world professional work, Elo): Haiku 5.5: 1620 | GPT-6 Luna: 1437 AA-Briefcase v1.1 (agentic knowledge work, Elo): Haiku 5.5: 1578 | GPT-6 Luna: 1336 OSWorld 2.1, offline subset (computer use): Haiku 5.5: 72.4% | GPT-6 Luna: 48.9% Humanity's Last Exam (no tools): Haiku 5.5: 45.9% | Luna: not reported Terminal-Bench 4.0 (agentic coding): Haiku 5.5: 39.2% | GPT-6 Luna: 16.4% FrontierCode 1.1 Main: Haiku 5.5: 46.4% | GPT-6 Luna: 42.4% Haiku 5.5 wins every published category, several by wide margins — Terminal-Bench by 23 points, OSWorld by 23.5. Two caveats before crowning it. First, these are vendor-published numbers; Haiku 5.5's scores span a wide range across its five effort settings (Artificial Analysis rates the Low variant at 29 on its Intelligence Index and the Max variant at 43). Second, these benchmarks run at short prompt lengths — they don't test the pricing cliff, which is where Luna fights back.

Where GLM-5.3 Flash Fits

GLM-5.3 Flash wasn't in Anthropic's comparison, but it holds its place in this band. On DeepSWE v1.1 it scored 63.4 — a 17-point post-training jump over GLM-5.2 on the same benchmark. On Z.ai's GDPval-AA v2 agent benchmark it sits at 1773 Elo, in the same range as Haiku 5.5's 1620 on v2.1 (different benchmark versions — treat as directional, not exact). On Artificial Analysis's Intelligence Index it scored 57, on an earlier index version than the 43 posted for Haiku 5.5 Max — again, directional only. What GLM-5.3 Flash has that neither rival offers: native video input and MIT-licensed open weights. If you need to feed screen recordings into the model, or you want the option to self-host, it is the only one of the three that can.

Where Each Model Wins

Use Claude Haiku 5.5 when prompts stay under 100K tokens and volume is high: subagents, classification, extraction, quick edits, summaries, support automation, and browser or computer-use loops. At $0.10/$0.50 with $0.01 cache reads it is the cheapest frontier-family model per task, and it is designed to pair with Opus 5.5 or Sonnet 5.5 as the fast worker underneath them — Cognition runs it as the sidekick in Devin Fusion with Opus 5.5 as the lead. Use GPT-6 Luna when context runs long or the task needs OpenAI tooling. Its $0.20/$1.20 rate is flat to 1.05M tokens — past 100K prompt tokens it undercuts Haiku 5.5 by 2.5× on input and roughly 2× on output. Hosted shell, computer use, MCP, apply patch, and reasoning-effort tuning remain the strongest agentic tool suite of the three. Use GLM-5.3 Flash for multimodal input at scale — images and video in the same context — and anywhere open weights or self-hosting matter. Strong coding scores for the price class, and the MIT license removes licensing ambiguity for commercial products.

The 100K Cliff Changes the Routing Math

Haiku 5.5's two-tier pricing is the detail most coverage will miss. Around 90% of requests to Haiku 4.5 fell under 100K tokens, which is why Anthropic could price that band so aggressively. But coding agents routinely exceed it: a full-repository context, a long debugging session, or a multi-file refactor blows past 100K quickly. Past that line Haiku 5.5 costs $0.50/$2.50 — more than Luna across the board. The practical rule: short-prompt, high-volume work goes to Haiku 5.5; long-context work goes to Luna. On Multos, Auto Router makes context-aware routing decisions per call — exactly the split this pricing structure rewards.

Effective Cost at Production Cache Rates

At an 85% cache hit rate and a 95:5 input/output ratio — the production averages Multos validates against, with prompts under 100K: Claude Haiku 5.5 blended: ~$0.05/M — at the floor, alongside DeepSeek V4 Flash GLM-5.3 Flash blended: ~$0.07/M GPT-6 Luna blended: ~$0.11/M For a 70M-token month (Starter allocation): Haiku 5.5 ~$3.30, GLM-5.3 Flash ~$4.90, GPT-6 Luna ~$7.40. The gaps are real but narrow at plan scale — task fit and context length should drive the routing decision before per-token price.

The One-Line Routing Rule

Haiku 5.5 for short-prompt volume. Luna for long context and OpenAI tooling. GLM-5.3 Flash for video input and open weights. Auto mode on Multos routes across all three — manual selection is for when you know the task shape upfront. Run all three against your own evals; these benchmarks tell you the ceiling, your workload tells you the floor.

Ready to Get Started?

All three models are on Multos. Pick from the model dropdown or let Auto Router decide.

Start Building Free