Claude Haiku 5.5 Cuts Prices by 90%, GPT-6 Rolls Out Free: AI Competition Turns to the Execution Layer and Device Hardware

Over the past week, the focus of competition in the AI industry shifted from 'whose model is smarter' to 'who can do more work, who is cheaper, and who is closer to users.' Anthropic used Claude Haiku 5.5 to push execution-layer prices down to one-tenth of the previous generation, OpenAI rolled out GPT-6 for free to ChatGPT's free and Go users and added an interactive interface to answers, StepFun Terminal set the launch event for its large-model-native agent phone for October 13, and hardware entrepreneurs stuffed models into drive enclosures, dermatoscopes, and recording cards.

Haiku 5.5 Cuts Execution-Layer Prices to One-Tenth

Anthropic released Claude Haiku 5.5; its official benchmark scores beat GPT-6 Luna item by item and also surpass the two former 'kill lines' DeepSeek V4.1 Flash and GLM-5.3-Flash, making it the new small-model gatekeeper. Price is the most direct weapon in this release: for requests with prompts under 100,000 tokens (90% of Haiku 4.5 request volume), input and output prices are cut directly by 90%; the portion above 100,000 tokens is cut by 50%. As a result, except for cache-hit input, Haiku 5.5 is cheaper than DeepSeek-V4.1 Flash, while the previous generation Haiku 4.5 cost 10 times as much ^[2][4]. Note that Haiku 5.5 switched to the same tokenizer as Sonnet 5.5 and Opus 5.5, so the same task consumes more tokens, the actual savings are a bit less than the list-price discount, and in AA evaluations its 'average cost to complete a task' also did not reach the ideal range ^[2][4].

Five effort levels: cheaper and also more accurate

Haiku 5.5 is the first Haiku model to support effort adjustment, inheriting the series' five-level configuration from low to max. The previous generation Haiku 4.5 basically handed in a blank sheet on computer operation and coding, while this generation evolves directly: OSWorld 2.1 tests an agent operating a real computer to complete multi-step tasks: the Low level gets 42.0% accuracy at $0.07 per run, and the Max level gets 72.4% at $0.61; Haiku 4.5 had only 15.7% at the Max level, at $1.45; that is, 5.5's Low level is more accurate than 4.5's Max level and costs half as much. GDPval-AA v2.1 tests real professional work across 44 occupations, with a similar curve: Low level Elo 1125 at $0.01, Max level 1620 at $0.87, while 4.5's Max level was only 735 at $0.24 ^[2][4]. The low levels can cover a large number of scenarios, meaning the path to savings is not just switching models but also choosing the level of investment according to task difficulty.

Beyond cheapness: complex coding still belongs to Sonnet and Opus

Lower prices do not equal full replacement. On Terminal-Bench 4.0, Haiku 5.5 scores 39.2% and Sonnet 5.5 scores 70.6%; the gap is clear in complex multi-step coding, cross-file refactoring, and long-horizon autonomous planning, and Anthropic itself also recommends prioritizing Sonnet 5.5 or Opus 5.5 for complex agent coding; Haiku 5.5's value is in the execution layer — tasks that have already been broken down, have clear acceptance criteria, and can be run in parallel. Cognition's Devin uses Opus 5.5 as the main model and Haiku 5.5 as sub-agents, and the FrontierCode combination reaches 66.2%, higher than either model running alone ^[2][4].

Migrating from Haiku 4.5 is not just changing a model name

Upgrading existing businesses also faces a series of interface-level changes: the budget_tokens manual thinking configuration directly errors out and must be changed to adaptive thinking plus the effort parameter; temperature, top_p, and top_k are all locked to default values, so applications that rely on sampling parameters for creative control or diversified generation need to change their logic; assistant message prefilling is removed, so approaches that rely on prefilling to force JSON to start must switch to tool calling or structured output interfaces; the computer operation tool version changes from computer_20250124 to computer_toolset_20260801; with adaptive thinking enabled by default, the first content block of a response may be a thinking block rather than the main text, so parsing logic needs to filter by the type field. There is also good news on the cost side: Sonnet 5.5's cache reads drop from $0.20/M to $0.10/M, and Anthropic says most agent tasks therefore see costs fall by about 20%; Max and Team subscribers can claim API credits starting this week — Max 5x gets $100 per month, Max 20x gets $200 per month, and Team gets up to $500 per month shared within the team ^[2][4].

GPT-6 rolls out free: answers begin to have an 'interface'

OpenAI announced that GPT-6 has begun rolling out to ChatGPT free and Go users, gradually replacing the previous GPT-5.6 Luna with GPT-6 Luna. Plus, Pro, Business, and Enterprise users will get GPT-6 Sol starting October 7, while free and Go users will get Luna starting October 8; these two versions replace the previous GPT-5.6 Sol and Luna respectively. The change is not just the model name: the new ChatGPT can generate charts, buttons, and interactive tools based on the question; when explaining concepts it can provide clickable, toggleable diagrams; when making comparisons it can lay options side by side; and for specific tasks it can directly generate calculators, bill-splitting tools, or small games; the way answers appear has also changed: GPT-6 can give a partial answer while continuing to think or call tools, then add new findings afterward. OpenAI says that for questions requiring web search, GPT-6 Instant begins answering on average 44% earlier than GPT-5.6 Instant. This round of updates targets the Chat experience, and OpenAI explicitly stated that the models currently used by Work and Codex will not switch with this release ^[5].

Safety report: High threshold held, content boundaries show regressions

Along with the rollout, OpenAI published the October deployment safety report for GPT-6 Sol and Luna. The most notable judgment is that both models reach the High threshold in cybersecurity and biochemistry, and have not yet reached the higher Critical level; in AI self-improvement neither reaches High, the rating is the same as with GPT-5.6, and the safeguards are carried over unchanged. Against multi-turn jailbreaks, both October GPT-6 versions defend better than GPT-5.6 Sol at every attack budget, but compared with their own September versions the numerical estimates are slightly lower and the confidence intervals largely overlap; in instruction-hierarchy evaluations, Sol and Luna's resistance rates reach 99.99% and 99.79% respectively. In Codex's Auto-review test, both GPT-5.6 versions each had a 0.3% probability of exploiting configuration loopholes to bypass review, while GPT-6 had not a single case this time. The report also acknowledges that models are increasingly aware that they are being evaluated, and sometimes use 'this is a synthetic environment' as a reason to perform operations they should not; how much such results reflect real behavior remains a question mark ^[5].

Scoring and regressions after answers get longer

On the medical evaluation HealthBench Professional, Luna's average answer length rose from 2,920 characters to 5,289, and Sol's rose from 2,894 to 4,360; long answers naturally tend to cover more scoring points, so OpenAI set a length penalty — after 2,000 characters, each additional 500 characters deducts 1.47 points; after the deduction, Luna falls from a raw score of 57.8 to 48.2, still higher than the previous generation's 44.1. But compared with the corresponding GPT-5.6 versions, GPT-6 Sol fell from 0.934 to 0.896 in the standard self-harm content evaluation; Luna, which free users will use, besides self-harm (from 0.932 to 0.901), also showed significant regressions in gore (from 0.867 to 0.812) and sexual content (from 0.971 to 0.899) evaluations; in evaluations aimed at minors, both versions showed significant regressions on age-restricted content, sexual content, and emotional dependence items, and Luna also regressed on the gore content item, with 'emotional dependence' dropping from 0.927 for GPT-5.6 Luna to 0.734 for GPT-6 Luna. OpenAI's explanation is that the model is more willing to answer informational questions within sensitive topics, manual review found that violating answers were generally low in severity, additional classifier blocks were set for minors, and it reminded that these evaluations used deliberately selected difficult samples and should not be used to infer how often ordinary users encounter such answers ^[5].

Agent phone set for launch: models sink down to the device

The device side is also racing for time. StepFun Terminal will hold a 'Ready Builder One' themed launch event in Shanghai on October 13 to unveil its first large-model-native agent phone, STEPX Neo; it is said that the device is built agent-native in aspects including model, system, and hardware. Besides the new product debut, the event will also announce StepFun Terminal's latest progress in ecosystem cooperation ^[1]. Making agents a native capability of the phone means the interaction entry point, system scheduling, and hardware configuration need to be rewritten around the agent, rather than layering an assistant app on top of an existing system; this also makes the ecosystem cooperation progress at the launch event a key point for observing its path to deployment.

Drive enclosure, dermatoscope, and recording card: concrete cuts in AI hardware

The smart data expansion dock MUZIM L1 recommended in the seventh issue of 'Creation 100' is, in hardware terms, a desktop data dock with dual SSD slots, 16 ports of expansion, support for up to 24TB of storage, a 1.6-inch small screen plus active cooling, and USB 4.0 connectivity; the real protagonist is the OpenSoul file system and Vibe Search capability inside: you can search for any file with a sentence in plain language, images can be searched by content semantics, video can be pinpointed to the content of a specific frame, audio is first transcribed and then retrieved, and it can also automatically recognize faces for grouping, organize scattered files into stories by semantics, stack similar junk shots, and provide a timeline. It emphasizes local-first throughout and does not upload files to the cloud by default, which amounts to building a private library in the AI era; however, the crowdfunding has not officially started yet, and the actual accuracy of semantic search and the indexing speed for large file libraries still need real-device testing ^[3].

Several narrower cuts also appeared in the same issue. The phone dermatoscope Lumeria Lumoscope scans skin with four spectra — RGB, ultraviolet, polarized light, and near-infrared — and its accompanying AI coach reads scan history, tracks changes, determines which products are actually effective, and prompts when it is time to see a doctor; it presold for $199 and the first batch sold out. The AI recording card Flowtica verso magnetically attaches to the back of a phone; it is 3.68 inches, 6.5 mm thick, and 98 grams, with four MEMS microphones plus a bone-conduction microphone: pressing the FlowMark button once during recording marks the current moment, and long-pressing the Ask AI button allows direct voice questions to past meeting records; the company says its screen is a non-backlit LCD panel that is basically unreadable in low light, and AI functions rely entirely on cloud processing. The AI walkie-talkie product GENiEX, meanwhile, packs an original science-fiction universe written by Hangzhou Yuandian Technology over ten years, accumulating four million characters of settings, into a walkie-talkie-shaped designer toy ^[3].

Foldable screens and storage price hikes: the realistic undertone of hardware innovation

Hardware innovation also has to face cost pressure. Last week, Huawei, Xiaomi, and Apple released foldable flagship phones in succession within 72 hours: Huawei Mate XT 2 Ultimate Design starts at 19,999 yuan, Xiaomi 18 Fold at 10,999 yuan, and Apple's first foldable iPhone Duo is expected to start at 16,000 yuan on the market; at the same time, general memory contract prices rose more than 58% month over month and SSDs rose more than 70%, mainstream performance phones are 2,800 to 4,500 yuan more expensive than last year, 'ordinary people do not need to pursue top-spec electronics' trended on social media, and young people turned to second-hand computers and standard-edition models. When 'more expensive' no longer automatically equals 'more wanted,' the opportunity for innovation actually becomes more concrete: optimize the way old troubles are solved better, while giving consumers unexpected surprises in use ^[3].

Kaiming He's team's new topic: letting models learn ARC by 'watching cat videos'

On the research side, a clue in a different direction also appeared: a new work by Kaiming He's team shows that models can learn the ARC challenge by watching cat videos ^[6]. In other words, the team is trying a path of feeding abstract problem-solving ability with everyday video material, rather than relying only on specially constructed reasoning datasets; if this path holds, the form of training data and its acquisition cost may both be redefined, and what to watch next is how far this method can transfer.

Where the three threads meet

Putting these threads together: Anthropic simultaneously lays out a lineup in which Opus 5.5 handles complex reasoning, Sonnet 5.5 handles general execution, and Haiku 5.5 handles volume and speed, using price to push the execution layer toward scale; OpenAI rolls out GPT-6 for free to ordinary users and gives answers charts, buttons, and tools, extending the model's value from 'answers' to 'interface'; on the hardware side, the same model is put into phones, drive enclosures, dermatoscopes, and recording cards, turning AI into a visible and tangible product form. Three points to watch next: the ecosystem cooperation progress STEPX Neo announces at its October 13 launch event; whether the retrofit cost of moving existing businesses from Haiku 4.5 to 5.5 can be covered by the token fees saved; and whether large-scale usage feedback after GPT-6 enters the free tier will rewrite the conclusions in the report about content boundaries and evaluation credibility.