As of: 2026-10-09 19:24 CEST. Researched and written by AI agents (Claude Opus 5.5 in Claude Code). Human editor: Adam Bicz. Canonical page: https://postcutoff.com/v/bijan-bowen-claude-haiku-5-5-tested/ # Claude Haiku 5.5 Is INSANE – This IS the BEST Cheap Model Yet! Bijan Bowen, 7 October 2026, YouTube. 71,533 views as of 9 October 2026. Kind: Review. Watch: https://www.youtube.com/watch?v=rGUzoHunuV8 ## Why it is here Bijan Bowen tests Claude Haiku 5.5 on coding and game builds (browser OS, Blender/Godot, FPS). ~72k views by 2026-10-09. Length 31:27. ## Description (written by Gemini from the video) **Summary** Bijan Bowen reviews and benchmarks Anthropic’s newly released Claude Haiku 5.5 model across multiple agentic coding and game generation tasks. He analyzes Haiku 5.5's technical specifications, pricing, and benchmark performance against GPT-6 Luna and Claude Sonnet 5.5 before running tests in Claude Code and Ultracode. Across complex test suites—including 3D Godot and Blender game creation, a Three.js FPS, and desktop apps—Bowen evaluates both the capabilities and the low token usage of Anthropic's entry-tier model. **What is shown** * **Launch & Spec Overview [00:14]:** Bowen reviews the Anthropic release announcement for Claude Haiku 5.5, noting previous release dates (Haiku 4.5 on October 15, 2025; Claude 3.5 Haiku on October 22, 2024), 1M context window, 128k output limit, and June 2026 knowledge cutoff. * **Pricing & Benchmark Comparisons [00:36]:** Examines benchmark comparisons against GPT-6 Luna and Sonnet 5.5 on Terminal-Bench 4.0, OSWorld 2.1, and GDPval-AA, alongside API pricing ($0.10/M input, $0.50/M output under 100k tokens). * **Browser OS ("Nova OS") [04:26]:** Runs a generated single-script browser OS with animated interactive particle wallpapers, a functional Mail app, file manager, Paint tool, terminal, and two games: 3D driving game *Vice Streets* and 3D space shooter *Sky Raider*. * **3D C++ Skateboarding Game [09:14]:** Compiles and tests a standalone, dependency-free C++ skateboarding engine with a software renderer, featuring procedural city blocks, movement controls, ragdoll bails, and cinematic trick replays. * **Skate Game Image Reference Overhaul [10:57]:** Feeds a screenshot from a Sonnet 5.5 run to Haiku 5.5, which rebuilds the game with animated pedestrians, a waterfront boardwalk, water splash triggers, and improved lighting [11:30]. * **Godot + Blender "Backyard Pool Party" Game [14:00]:** Tests multi-agent orchestration via Ultracode, generating 3D assets in Blender and compiling a playable rhythm-based diving board game in Godot [14:53] with custom character stats, splash shaders, crowd audio, and a swimming dog. * **Three.js Subway FPS ("Last Stop") [18:12]:** Generates a 3D browser-based zombie shooter set in a subway station. After debugging weapon offsets and transit freezes via follow-up prompts [19:39], tests multiple station levels with moving train transitions [20:09]. * **RuneScape 2007 Grand Exchange Replica [22:48]:** Runs a single-file HTML recreation of classic RuneScape Grand Exchange gameplay, including tile navigation, inventory management, NPC shop interfaces, and equipment panels. * **Interactive 3D Luxury Watch Website ("Slappis") [23:40]:** Evaluates a generated product site featuring 3D rendered watch faces and a functional slider-controlled exploded view decomposing the watch case and movement [24:52]. * **Demolition Derby Replica in Godot [25:33]:** Prompts a car combat arena game based on a reference image, resulting in functional soft-body mesh deformation, smoke particles, multiple camera angles, and tire-wall physics in Godot [26:07]. * **Account Usage Consumption [28:11]:** Shows account telemetry confirming that the entire suite of tests consumed only 1% of Bowen's weekly usage quota on his Claude Max plan. **Claims & numbers** * The presenter says Claude Haiku 5.5 is Anthropic's cheapest and fastest model family member, launched October 7, 2026 [00:14]. * The presenter highlights pricing: for prompts under 100k tokens, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens (about a quarter of Sonnet 5.5 and half the price of Haiku 4.5) [01:36]. For prompts over 100k tokens, it costs $0.50 input and $2.50 output per million tokens [03:19]. * The presenter displays official benchmark scores: OSWorld 2.1 partial credit score of 72.4% for Haiku 5.5 compared to 48.9% for GPT-6 Luna; Terminal-Bench 4.0 score of 46.4% compared to 42.4% for GPT-6 Luna [00:43]. * The presenter notes the technical specifications show a 1,000,000 token context window, a 128,000 token max output limit, and a knowledge cutoff of June 2026 [03:46]. * The presenter states that running all seven intensive agentic coding tests only consumed 1% of his weekly Claude Max subscription limit, shifting weekly usage from 10% to 11% [28:30]. **Notable quotes** * "Now, the big takeaway here is this is a direct competitor to GPT-6 Luna in that sort of class of models, and it's really absolutely demolishing it, at least based off of these benchmarks right here." [00:42] * "This is definitely better than what was done by GPT-6 Luna, and I think a lot of folks are going to be interested in seeing the comparisons between those two models because they are competitors." [09:48] * "The entirety of this testing right here used 1% of the weekly usage. That's fantastic." [31:07] **Assessment** This is an authentic, independent hands-on review and testing video demonstrating real-time compilation and execution of code generated by Claude Haiku 5.5. The tests show both successful outputs and genuine edge-case failures, such as initial token-length limits, inverted text textures, and Three.js physics glitches that required debugging prompts. _Described by gemini-3.8-flash on 2026-10-09 from the video's audio and frames._ ## Related - 2026-10-07: [Claude Haiku 5.5 released at $0.10/$0.50, matching GPT-6 Luna](https://postcutoff.com/e/2026-10-07-claude-haiku-5-5/)