Haiku 5.5 vs GPT-6 Luna: Which Cheap AI Is Worth Using?

Original author：Aditya Bharti

Original source：https://www.youtube.com/embed/tWjqX3K9WeI

Claude Haiku 5.5 vs GPT-6 Luna: I gave both the same invoice extraction, support-routing and Python refund-rule tasks. Then I compared the published token prices, context limits, caching and processing options behind “cheap intelligence.” Both passed these narrow task groups. The listed short-prompt API rates match, but longer prompts can produce a different cost comparison. The video shows a worked 150K-input-token example and a 10,000-request budget, including the effect of additional reasoning tokens. These costs use stated assumptions, not measured API bills. CHAPTERS 00:00 Can cheap AI do useful work? 00:28 Test setup and model versions 00:54 Invoice extraction and missing fields 02:07 Support routing and human review 02:52 Python refund rules and boundary tests 03:40 GPT-6 Luna results 04:29 Input and output token prices 04:52 Context and output limits 05:14 The long-prompt cost gap 05:47 Caching: reads, writes and misses 06:15 Batch, Flex and Fast pricing 06:39 Worked budget and reasoning costs 07:08 Choosing by cost per accepted result TEST SCOPE One synthetic prompt per product: Haiku 5.5 Medium in Claude web and the October GPT-6 Luna experience in ChatGPT Go at its default setting. Product/reasoning settings were not matched. October ChatGPT Luna differs from the API/Codex versions. Both generated functions passed the same 12 predefined local checks. This does not establish an overall winner, general accuracy, prompt-injection safety, or production reliability. We did not measure API token usage, latency or throughput. Invoice case identifiers were excluded from scoring because the original key was ambiguous; field values were checked consistently. PRICE BASIS USD first-party API rates checked 9 October 2026. Equal assumed token counts do not imply equal tokenization of the same text. Billed output includes reasoning. Cache eligibility, tool fees, retries, regional settings, provider pricing and review can change total cost. Chat subscriptions and API billing are separate. PRIMARY SOURCES Anthropic launch: https://www.anthropic.com/claude-haiku-5-5 Haiku specs: https://platform.claude.com/docs/en/models/haiku-5-5/overview Anthropic pricing: https://platform.claude.com/docs/en/about-claude/pricing GPT-6 Luna API: https://developers.openai.com/api/docs/models/gpt-6-luna October ChatGPT rollout: https://openai.com/index/gpt-6-for-everyone/ Model-version distinction: https://deploymentsafety.openai.com/gpt-6-october/model-data-and-training Reasoning billing: https://developers.openai.com/api/docs/guides/reasoning Claude thinking: https://platform.claude.com/docs/en/build-with-claude/thinking Made by Aditya Bharti. Synthetic examples are labeled. Narration uses my locally cloned Voicebox voice; explanatory animation and thumbnail artwork are AI-assisted. No paid promotion. #ClaudeHaiku #GPT6 #AI