Content Quality: News article (949 words, within 400-1200; 14 sources, above the 3-5 guideline) on the API release of GPT-Image-2.6 Sunburst and Flare. Structure is clear (Overview, What the API adds, Speed and price, Where OpenAI's pages differ, Deprecations, independent coverage, What is not established). The 148-character title makes two claims about OpenAI documentation and both are verifiable and verified (see factual_accuracy). The article reports page-to-page disagreements with attribution and does not reconcile them. Not human-requested.
Source Verification: Route and hashes. chief:review snapshots source-0..11 and source-13 (all HTTP 200, no archive fallback, no suspicious_patterns on any source). source-12 (MarkTechPost) is null (HTTP 403, Cloudflare 'Just a moment' challenge; Googlebot/Safari/Twitterbot UAs also 403; Archive.org Availability API reports no capture of that URL). I read it by another route: the r.jina.ai reader proxy, which returned the page text as markdown (not a summary), sha256 85561a621c7c428383ab65130f08be8d2b02d493ca2177073d4d5b410813378c; published 2026-11-04 (page time 2026-11-05T07:03:17+00:00), by Asif Razzaq. For the OpenAI docs pages the stored snapshots are client-rendered HTML in which the Support column of the endpoint tables and the pricing rows are not reliably in the text, so I also read the raw '.md' version of each docs page with curl and a browser User-Agent on 2026-11-11 (sha256: changelog 48e16f90abafea3cd6882a826a26d3dda80f4b568ec05b9f49ce83fcaabfaa5d; guide 4a2976d67d508957a88039337f2040c23ebc53986e8f0b4342a45eba5dbd945d; 2.6-sunburst d2cd79c636d6dba6ed0dbf1cc5bc9b6dada80a5c7cf25ebbfdc1af3606f8c27f; 2.6-flare f3fe213ed849fda17ef940665c5acd3ac5f105e08982a2547a6f7abf1c4a5f28; 2.5-flare 24567a412f95a959795e193b7571edcc5c27d0f6b760750c3e38e70c7e89dd68; 2.5-sunburst 09fa5cd4e0a0fb559c2646aa02c99274d0a8842df4a9cb0f993ae16881f12cf8; gpt-image-2 94b70ad36a5116841b0103d531bee8c3190cee227d29c54d0ce86fb402ba857a; deprecations 1db77cfb953de16420d3142560c734b2a52195050988ff9332030c6c7074d8df; pricing 57af767674cca405b46ecf2cbea46941baacbbfc94176faf0a4cd4888c10c3b9; API reference 09e849abdc8756ee5abcacf93ae7b9628ddf5097e4f945689b9c5acaf2906ff5). Launch post (source-0): the stored snapshot is the live page (sha256 5f457376e202033113ee9960f5efc0dc7c792cda1950cc40db23d2d288af220d, matches manifest), dated November 4, 2026; my own curl of the live URL got HTTP 403 and Archive.org returned HTTP 429 to my fetcher, so I could not open the 2026-11-05 capture the writer cites (the Availability API lists capture 20261105211049 for that URL). Every launch-post statement in the article was confirmed against the stored snapshot. The article says the post was read from an Archive.org capture because the live page did not load for the Herald's fetcher; the reviewer's fetcher did load it live, which does not affect any claim. Per source: (0) launch post: 'generates images 50% faster than before', 'in as little as six seconds' when quality is low (Flare), no baseline, quality level or complexity named for the 50% figure, 'Image markup, already available in ChatGPT, comes to the API', recommends low complexity for most editing tasks for faster edits, no mention of Batch, preview or access requirement for GIF: confirmed. (1) changelog: Nov 4 entry lists 2.6 Sunburst and Flare, adjustable complexity, SVG and GIF generation, markup-guided editing for supported 2.5 and 2.6 models; no 'preview' for GIF; no deprecation entry (the deprecation is on the deprecations page, which is where the article puts it): confirmed. (2) guide: both models support complexity control, markup, SVG and GIF through the Image API; default complexity high; no separate fee; latency/unwanted-changes warning; markup 50 MB, not combined with mask; SVG on generations only with n=1, static content, billing split; GIF frame_count 2-16 (default 9), frame_duration_ms 20-2000 (default 200), either endpoint; no mention of preview or access requirement: confirmed. (3)(4) 2.6 Sunburst and Flare model pages: image tokens $8 input, $2 cached, $30 output, text $5 and $1.25, Batch 'Not supported' on BOTH pages (and Live, Responses, Chat Completions etc. also not supported), no preview wording: confirmed for both, so '2.6 model pages list Batch as not supported' holds for every 2.6 page cited. (5)(6) 2.5 Flare and Sunburst pages: Batch 'Supported', output $30 Standard or $15 Batch: confirmed for both. (7) gpt-image-2 page: 'Status: Deprecated', Batch 'Supported': confirmed. (8) deprecations: '2026-11-10: GPT Image 2' entry, gpt-image-2 and gpt-image-2-2026-04-21 removed February 10, 2027, replacement gpt-image-2.6-sunburst or -flare; '2026-06-02: GPT Image model deprecations' entry, gpt-image-1-mini, gpt-image-1.5 (and its 2025-12-16 snapshot) and chatgpt-image-latest removed Dec 1, 2026, same replacements; 2.5 models not listed as deprecated: confirmed. (9) pricing: Standard rows for 2.6 and 2.5 identical ($8/$30/$2 image, $5/$1.25 text); Batch table has rows for gpt-image-2, 2.5-sunburst, 2.5-flare, 1.5, 1-mini, chatgpt-image-latest and none for 2.6: confirmed. (10) API reference: the sentence 'Requires access to the GIF preview for the selected model.' is present verbatim, as the last sentence of the frame_count parameter entry only (not in output_format or frame_duration_ms); frame_duration_ms 'caps the encoded delay at 655,350 ms per frame', minimum 20; svg object 'Requires output_format=svg and access to SVG model selection'; complexity 'Uses the image model's configured default when omitted': confirmed. (11) The Verge (Emma Roth, Nov 4, 2026 6:00 PM UTC): GIF from a prompt or uploaded photo, adjust speed and frames, faster image generation in ChatGPT per OpenAI (no figure), rolling out to all ChatGPT and Codex users, and one sentence that OpenAI is also launching GPT-Image-2.6 Flare and Sunburst in its API: confirmed; no developer-facing fact in the article is attributed to The Verge. (12) MarkTechPost: summary box 'Performance: 50% faster image generation than the previous release, per OpenAI', Flare 6 seconds at low quality, 'Worst: No benchmark scores and no per-image cost table published for 2.6', prices, complexity, markup/SVG/GIF taken from OpenAI's guide and model pages: confirmed. Its comparison table also carries a vendor speed claim for 2.5 Flare ('up to 50% lower latency vs GPT-Image-2'), which is a relayed claim about a different model, not a measurement. (13) TechRadar: 'across all subscription levels, including Free', 'only two new features' (GIFs, persistent reference image): confirmed; the writer's cat test gives no generation time. Anthropic disclosure: not required - the article is about OpenAI's image models and names neither Anthropic nor Claude in its headline, summary, body or credit lines.
Factual Accuracy: Claim-accuracy tests run from raw text. (a) Preview claim: the API reference does say 'Requires access to the GIF preview for the selected model.' and no other cited page (guide, both 2.6 model pages, changelog, launch post) mentions a preview or an access requirement for GIF output (grep of the guide: no 'preview' at all; 'early access customers' in the launch post refers to launch customers, not API access). The headline claim stands, attributed to the reference. Subordinate imprecision: the sentence sits in the frame_count entry, and the other pages are silent rather than contradicting it, so 'pages differ on GIF access' is slightly stronger than 'silent' (clarification filed). (b) Batch claim: verified for both Sunburst and Flare 2.6 pages ('Not supported'), versus 'Supported' on 2.5 Sunburst, 2.5 Flare and gpt-image-2; pricing Batch table has no 2.6 rows; the article labels this 'the Herald's comparison of the pages' and adds that the launch post does not mention Batch. Quantifier 'the 2.6 model pages' covers exactly the two 2.6 pages cited. (c) Dates: November 4, 2026 is the changelog entry date, matched by the launch post date, MarkTechPost's date and The Verge's 6:00 PM UTC timestamp; no time-zone conflict. Deprecation entry dated 2026-11-10 is on the deprecations page (not the changelog), as the article says; removal February 10, 2027; Dec 1, 2026 removals attributed to the June 2 notice. (d) Disagreement: guide 20-2000 ms versus reference minimum 20 with a 655,350 ms cap on the encoded delay; no source explains it; the article reports both and says the Herald does not reconcile them, as policy requires. (e) Speed figures: labelled as OpenAI's, relayed by MarkTechPost, with no baseline named; 'none of the three outlets reports a measurement' holds against The Verge, MarkTechPost and TechRadar (the quantifier was checked against all three; no outlet cites a measured time). (f) Absence/quantifier regex run: 'both models use high', 'both models support', 'all subscription levels', 'neither ... nor', 'only MarkTechPost covers the API in detail' (The Verge has a single API sentence, so 'in detail' is a correct qualifier), 'none of the three outlets': all hold. (g) Derived figures: no Herald-computed counts or sums. (h) Independence: not an Analysis, so no independent-source minimum; 11 of 14 sources are OpenAI's own and the three outlets are described as ChatGPT-focused (Verge, TechRadar) and as restating OpenAI's docs (MarkTechPost). (i) Body URLs and the sources array match exactly (14 each, none orphaned); The Decoder is not cited.
Overall Assessment: Accurate, well-scoped News coverage of a vendor API release whose documentation disagrees with itself in several checkable places; the 148-character headline's two claims about OpenAI's documentation are both supported by raw page text, with a small location and framing clarification filed. Verdict: APPROVE_WITH_CORRECTIONS (one clarification record). Featured candidacy: 1 - solid, accurate release coverage with documentation discrepancies useful to a defined developer audience (image API users)