Model Buzz

In the repliestwendah asked why Haiku 5.5 xhigh was missing; Laprablenia asked about Gemini 3.8 Flash. Right-Performance-93 preferred Sol medium based on the same author's table, without an independent run.

The author explicitly cautioned that provider allowance percentages are not equivalent compute units. Exact prompts, repeated runs, account windows and execution logs were not verified; the table does not establish general quality rankings or API costs.

Original October 9 post and replies pevpx95, pewm3lp and pf04ai5. Model names, usage figures and quality ratings are self-reported.

In the replieslgnorant-lu reported 2/5 using Gemini Web 3.8 Flash extended with rephrased spatial questions, questioning which blue and white books the bookshelf prompt meant. The author conceded that question was problematic. The changed prompts make this a different attempt.

Single attempts with different settings, clients and providers; these user-assigned grades do not establish a model ranking or reproduce the HSS paper.

LINUX DO posts 23449203, 23450507 and 23450965, October 9. Model labels, grades and client behavior are self-reported.

In the repliessensei_von_bonzai attributed the gap to Haiku’s higher prices above 100K prompt tokens; the run’s per-turn context was not verified. _JohnWisdom questioned voxel generation as a Haiku benchmark. DesperateSunday reported little weekly-allowance movement on five smaller tasks, a different self-reported workload.

Both model names and xhigh settings are self-reported. Anthropic publishes higher per-token prices for prompts over 100K, but that does not verify this run’s aggregate token totals or cost calculation.

Original post and human replies pej3rih, pej4ybz and pejywq4. Exact prompts, raw token traces, repeats and compaction behavior were not verified; no general price or quality ranking follows.

In the repliesThe HN API returned no comments at the snapshot; no reply-based assessment was available.

Vals reports these measurements; TechReadly did not reproduce them. LegalBench covered 12 public subtasks rather than the full benchmark, with possible training contamination. Latency tests used one residential machine and unequal concurrency. The HN submitter’s affiliation with Vals is unconfirmed.

HN submission 49984418 links Connor Frank’s October 6 Vals evaluation at https://www.vals.ai/blogs/independent-evaluation-of-jev. Jev returns fixed-choice probabilities; the findings do not establish general chatbot or legal-workflow superiority.

In the repliesPackageDecent3526 recommended Sonnet 5.5 for architectural intent and edge cases. Cyvster suggested Qwen, reporting that a local 9B model had found issues in DeepSeek-written code. These were individual recommendations rather than a comparison on the author’s repository.

Exact Kimi and Qwen versions, prompts, client, hardware and API plans were not provided. The reported fixes and code quality were not independently tested.

Original October 5 post and human replies pe22fk4 and pe6vsd0; self-reported model names and experience. No general model ranking follows.

In the repliesThe comment listing was empty when checked; no reply-based confirmation or challenge was available.

These are one author’s small-task results, not a complete model ranking or an independently reproduced benchmark. Newly generated tasks do not establish freedom from contamination.

Original post reviewed October 6. Strata is linked in the post; the author’s relationship to the project is unconfirmed. Raw execution traces were not inspected.

In the repliesevilncarnate82 described manually handing Claude’s architect prompts to Antigravity and adding canary tests after incomplete work. MentalThroat7733 reported using AI Studio and Antigravity with Android Studio, or copying between Gemini and an editor for web/Python work, mainly with 3.8 Flash extended thinking and automated tests. These were different workflows, rather than tests of bocondo’s ticket.

The author’s exact Gemini version, prompts, settings and issue severity are unknown. One self-reported review comparison does not establish model accuracy or a general ranking.

Original post and substantive replies; subscription entitlements and CLI-retirement claims in comments were not verified and are not adopted here.

In the repliesOne reply asked for the exact prompt before judging the outputs. hurdurdur7 requested steps and resolution. Ipwnurface said camera movement was achievable without a LoRA in their own setup, challenging a universal camera-angle limitation.

The model labels come from the post. Provider, exact prompt, resolution, and generation settings were not verified; one workflow does not establish model rankings or official prices.

Original post and replies pdehzf5, pdg1tv5, pdl28f4, and the author’s pdfaigk qualification; posted October 2 and reviewed October 6.

In the repliesCommenter ricedumpling described a separate coding workflow: use GPT Plus to develop a plan, save business logic and interface details as Markdown in the project, then have Claude Code connected to the official DeepSeek API read that plan and execute. When zzbd questioned the cache cost of switching models, ricedumpling clarified that GPT served as a consultant while execution and input caching stayed on DeepSeek; no measured savings were supplied.

The author did not identify a DeepSeek version. Evening use is a personal routine, not a verified discount schedule.

Original post and named replies on LINUX DO; spending and workflow results are self-reported, not independently tested.

In the repliesComrade-Porcupine also said GPT-6.1 Sol quickly exceeded its context and required compaction, and said they planned to use Astra for design and GPT-5.6 Sol for implementation. Automatic_Cookie42 called it an improvement over Sol 6 without describing a task or test. This is one anecdotal run: the thread provides no prompt, configuration, logs or independent reproduction. It is not a benchmark.

OpenAI’s September 29 launch page presents GPT-6.1 Sol as an upgrade to GPT-6 Sol. The company’s product claims do not verify this commenter’s experience.

One Reddit thread reviewed September 30, 2026. The performance comparison is a commenter’s single self-reported run without published logs or independent reproduction.