Two sweeps in one filing: a live agent-readiness pass on pointcast.xyz, and a sourced outlook on the model landscape for September 2026 through February 2027.
PART I · The surfaces are excellent and honest; the clock is the problem. Every one of the 377 URLs the manifest declares is live, 50 MCP tools answer cold, and the x402 handshake is spec-correct — but the site hasn't shipped in 9 days, the self-inspector is 36 days stale against its own promise, and the flagship write path is dead on one unbound env var.
PART II · Agents got cheap, video got aggregated, music got sued, and the browser war ended in Chrome. Expect a loud October–December (Gemini 4, Grok 5, Amazon's re:Invent model, maybe a Fable 5.x), open weights ~4 months off the frontier with the coding gap closed, and the first hard deadline 24 days out: the Sora API dies September 24.
Method: read-only probes against the live site on 2026-08-31 — every declared endpoint status-checked, MCP handshakes run, x402 402 bodies decoded, six pages parsed for structured data, all cross-graded against where the agent web actually stands this month.
All 400 unique URLs in agents.json were probed. Every literally-declared URL — 377 of them — returned 200. Every pattern endpoint resolved once fed a real value. 20/20 sampled JSON bodies parse. A junk URL returns a genuine 404, so the 200s are real content, not an SPA catch-all.
All 96 internal links in /llms.txt checked exhaustively: 96/96 live. robots.txt names 14 AI crawlers with explicit Allow and its three Sitemap lines all point at files that exist. The homepage <head> is a model of in-band discovery — rel=alternate to the manifest, both llms files, feeds, and ten channel RSS feeds.
Both endpoints speak protocol 2025-06-18, statelessly — tools/list and tools/call succeed cold with no session id. 50 tools, identical on both, with self-documenting instructions. CORS preflight on /api/mcp-v2 is textbook. The self-describing GET pattern on write endpoints (/api/ping teaches you how to POST to it) is better than most of the 2026 agent web.
The 402 carries a proper x402 v2 accepts array — $0.01 USDC on Etherlink via Permit2, facilitator header exposed through CORS, and a clean 400 on malformed signatures. Still a proof-of-concept selling only its own receipt, but the plumbing is right.
Nothing here is availability — it's all coherence and freshness, which for a site whose brand is a transparent ledger is the part that counts.
/health.json says "inspections run at deploy time," but inspectedAt is 2026-07-26 while the site deployed 2026-08-22. The public "clean, 261 doors open" verdict predates five weeks of changes.
Relative timestamps were baked at build time on Aug 22. An agent reading the HTML today is told a PR merged three minutes ago. Later items in the same list use absolute dates — inconsistent and, right now, false.
The endpoint self-admits it: until PC_PUBLISH_KV is bound in the Pages env, POST returns 503. The publishing queue advertised since v1 is a door painted on a wall. (/api/talk is likewise an honest stub — it returns mock success.)
Every UA stanza carries Disallow: /api/, yet llms.txt and agents.json actively promote /api/mcp, /api/mcp-v2, and /api/ping. Robots-respecting agents are formally told to stay away from the endpoints you built for them.
agents.json lists 10 channels / 10 block types; llms.txt says 9 and 8; /for-agents says 8 and 8. And /for-agents' visible dateline still reads 2025-01-14. An agent cross-checking prose against manifest sees drift.
Channel pages answer to slugs, but block pages display codes — /c/gdn.rss 404s while /c/garden.rss works. ?station=el-segundo 404s on a weather API about El Segundo (real slugs: santa-monica, hermosa…). Mood slugs, race slugs, and yee ids have no declared index at all; zen-cat wants YYYYMMDD where its sibling wants an integer. Every 404 an agent hits here is a template filled with a plausible-but-wrong value.
/sitemap.xml 404s at the conventional path (the three real sitemaps exist, but nothing redirects). .well-known/ai-plugin.json declares agents.json as an OpenAPI spec — parse failure for anything that believes it. POST to the x402 endpoint gets a bare 405 with no Allow header. GET on the MCP endpoint with an SSE Accept returns an HTML landing page. OPTIONS preflight on static JSON returns 405 despite headers promising otherwise.
The clock is the finding. Feeds, manifest, MCP's now_snapshot, and the wire all timestamp 2026-08-22 — nine days ago. Nothing distinguishes "quiet week" from "abandoned": no cadence field, no nextUpdate, no re-run inspector. For a living broadcast, freshness is the product claim, and right now every visible clock on the site undercuts it. The single highest-leverage change in this filing is making timestamps regenerate at deploy and letting the inspector actually run when it says it does.
Ordered by leverage. Everything above the line is under an hour each.
| # | Fix | Where |
|---|---|---|
| 01 | Bind PC_PUBLISH_KV — flips /api/publish from advertised-but-503 to liveSame family as the unset X402_RECEIPT_SK and YARD_RESIDENT_KEY: one dashboard visit closes three dead paths | CF Pages env |
| 02 | Wire the Town Inspector + agents.json regeneration into the deploy step so inspectedAt/generatedAt always equal deploy time | build pipeline |
| 03 | Audit Cloudflare AI Crawl Control before Sept 15 — the mixed-use-crawler default flip could silently block wanted traffic | CF dashboard |
| 04 | robots.txt: Allow /api/mcp*, /api/ping, /api/x402/receipt; add a Content Signals block stating search/ai-input/ai-train intent | robots.txt |
| 05 | 301 /sitemap.xml → sitemap-index.xml; add the discovery + blocks sitemaps as entries in the index; 301 /agent-readiness → /health | _redirects |
| 06 | Channel code→slug aliases (/c/gdn.* → /c/garden.*); alias station=el-segundo to home coordinates; enumerate valid stations in the 404 body | functions |
| 07 | /wire: absolute ISO timestamps (or <time datetime>) instead of baked "3m ago"; point its rel=alternate at feed.xml like every other page | template |
| 08 | Add valuesFrom hints to every pattern endpoint; ship /moods.json; declare the /yee index; note zenCat's YYYYMMDD format; declare method: POST on write endpoints | agents.json gen |
| 09 | Sync channel/type counts across llms.txt, /for-agents, agents.json from one source; fix the /for-agents dateline; fix or drop ai-plugin.json's fake OpenAPI block | generators |
| 10 | Fill CreativeWork JSON-LD (author, articleBody, image, url) on block pages from the data already in each block's JSON; put full body text in feed.json content_text | templates |
| 11 | Give POST-on-x402 a real 405 (Allow header + one-line pointer to the GET flow); content-negotiate SSE on MCP GET; fix OPTIONS on static JSON | functions |
| 12 | Submit the MCP server to the official registry, and submit sitemaps to Bing Webmaster — ChatGPT search rides Bing; this is the discovery lever that actually works | external |
The external benchmark matters because it says which of your bets are early versus wasted. As of this month:
Sourced sweep across six domains, adversarially fact-checked (20 load-bearing claims re-verified; corrections folded in). Chips mark confidence: confirmed announced rumored speculative
Every lab converged on the same axis: long-running agents that browse, code, and transact. Prices collapsed a full order of magnitude at the low end while the top tier held.
| Lab | Where it stands (late Aug 2026) |
|---|---|
| Anthropic | Fable 5 / Mythos 5 (June 9) is the frontier — Mythos-class sits above Opus, ~95% SWE-bench Verified, $10/$50. The June export-control episode took it offline globally for 19 days; it's been usage-credit-only on seat plans since June 23. Opus 5 (July 24, $5/$25, effort toggle) delivers near-Fable performance at half the price and is the default on Max. Claude's built-in Cowork browser began rolling out Aug 26; Claude in Chrome went GA the same week. |
| OpenAI | GPT-5.6 family (July 9): Sol / Terra / Luna. Luna at $0.20/$1.20 is the free-tier default; Sol promo-priced $4/$20 through ~Nov 21. Sora is dead (app Apr 26, API Sept 24); Atlas browser killed Aug 9, folding agentic browsing into the desktop app + a Chrome extension. A next family reportedly codenamed "Astra" targets long-horizon multi-agent work — no date, no card. |
| Gemini 3 era; 3.5 Pro delayed three times over coding shortfalls — Google shipped three Flash models instead (July 21) and confirmed Gemini 4 pretraining, agents and coding first. Project Astra powers Search Live. Prediction markets put ~85% on Gemini 4 by Nov 30. | |
| xAI | Monthly cadence off Colossus 2 (~1GW): Grok 4.6 (Aug 12, 500K context, agent-tuned); Grok 4.7 on a new 2.1T foundation expected ~early September; Grok 5 in training, slipping toward Q4. |
| Meta | Llama shelved as the frontier line. Muse Spark (closed, April) then a partial re-opening: Muse Glimmer (Aug 10) — 30B, Apache 2.0, distilled from Spark, runs on a 24GB GPU or a Mac. Muse Video and Spark 1.2 open weights stated ahead. |
| Amazon / Microsoft | Amazon axed most Nova models to bet on one frontier model (reportedly re:Invent), on ~1GW of Trainium; its New Carlisle campus for Anthropic hit gigawatt scale before Stargate. Microsoft's MAI line is real but Suleyman says frontier-class is 12–18 months out. |
Watch: announced Gemini 4 · announced Grok 5 · rumored OpenAI "Astra" family · rumored Fable 5.1 (community chatter only; nothing official) · rumored Amazon's re:Invent debut · speculative wider Mythos-tier trusted access. Base case: two to three frontier flagships land October–December.
The benchmark vacuum is real. OpenAI dropped SWE-bench Verified in February (59% of failed tasks had flawed tests), then retracted its SWE-bench Pro recommendation too (~30% flawed tasks, ~70% noise ceiling). There is currently no consensus coding benchmark; Terminal-Bench-style long-horizon evals are the direction, and whichever replacement wins becomes the fought-over scoreboard this winter. Discount any single headline number until then.
The summer story: Max-class Chinese open weights arrived, roughly four months behind the frontier, with the coding gap effectively closed. Licenses are where the traps are now.
| Model | What shipped | License |
|---|---|---|
| DeepSeek V4 | V4-Flash open weights + API July 31 (284B/13B active); V4-Pro GA as Pro-0813 on Aug 13 (1.6T/49B). 1M context, ~80.6% SWE-bench Verified at ~1/7 frontier output price. R2 still held back — the checkpoint floating around isn't official. | MIT |
| Kimi K3 | 2.8T sparse MoE (July 16; weights July 27) — largest open-weight model ever, 1M context, Agent Swarm lineage. Needs ~1.6TB of GPU memory to self-host. | customrev/attribution clauses |
| Qwen3.8 | First-ever open-weight Max-class Qwen (Aug 12): 2.4T-A95B, context up to 1M — plus the 27B sibling that local-AI writers call the most important local release of 2026. | custom (Max)27B family friendlier |
| GLM-5.2 / 5.3 | 5.2 (June, MIT, 753B-A40B) topped the open coding indexes; 5.3 (Aug 14, weights ~Aug 28) claims strongest open coding model. The sleeper: GLM-5.3-Flash, 320B-A18B, reported MIT — the actually self-hostable member. | MIT / custom |
| MiniMax M2.7/M3 | Cheap self-evolving agent line (~8% of Sonnet's cost); M2.7 partly trained by its own predecessor. | open (HF) |
| Gemma 4 | The on-device standard: E2B→31B, natively multimodal, 256K context, agent/tool-calling focus; E4B runs in ~5GB at 4-bit. | Apache 2.0 |
| Muse Glimmer | Meta's re-opening gesture: 30B dense multimodal agentic, under 20GB footprint. | Apache 2.0 |
| gpt-oss | Still just 120b/20b from Aug 2025; no successor announced. Now an MLPerf benchmark — infrastructure, not news. | Apache 2.0 |
Local stack: MLX won on Apple Silicon (Ollama quietly switched backends to it), and 27–31B is the sweet spot — Qwen3.8-27B, Gemma 4, Glimmer all run fast on the iMac. An always-on local resident agent for the broadcast, with no API bill, is now a real option.
Watch: rumored GLM-5.5 (1T+, long-horizon coding agents; missed its August window) · rumored DeepSeek R2 or a V4.5 · announced Mistral's frontier-gap model GA · speculative Kimi K3.5 (~Oct–Nov by cadence) · speculative gpt-oss-2 under competitive pressure.
The market picked aggregators over models. Sora is dead, Runway now routes competitors' models, and Higgsfield — already wired into the TikTok pipeline — runs ~$230–300M ARR bundling 50+ models. Native synchronized audio is table stakes across every current flagship.
| Model | State |
|---|---|
| Seedance 2.5ByteDance | Launched July 31; the developer API went live Aug 7 on Volcano Ark and BytePlus ModelArk. 30-second clips, up to 50 multimodal reference inputs, audio co-generated in the same latent space. One catch the press release glossed: shipped output tops out at 720p despite the native-4K announcement. |
| SoraOpenAI | App dead April 26; API terminates Sept 24. ~$1M/day compute against ~$2.1M total revenue. No successor product named. |
| Veo 3.1 / Gemini OmniGoogle | Veo 3.1 everywhere (native vertical, 48kHz synced audio); Flow passed 275M videos in five months. Gemini Omni Flash tops the no-audio arena; the full Omni model rolls toward Vertex and YouTube Shorts Remix — the biggest free on-ramp in the space. |
| Kling 3.0Kuaishou | Native 4K/60fps/15s with multilingual lip-sync; ~$240M ARR. Turbo variant since June. |
| Hailuo 3.0MiniMax | Multi-shot 2K with native stereo at $0.08/s; tops several arena boards. H3-Base weights are open but the community license excludes deployment in the US, EU, UK, and South Korea — not self-hostable from El Segundo. |
| Wan 3.0Alibaba | GA Aug 24 ($0.05–0.20/s, doc-to-video, leads the with-audio arena) — API-only. The open line stops at Wan 2.2; "Wan 3.0 open weights" claims circulating are fabricated. |
| Runway / Luma | Gen-4.5 + Aleph 2.0 editing; Runway's real move was launching a model router in July. Luma Ray3.2 does 16-keyframe control and native HDR. |
Cost curve: ~3x cheaper per generation cycle through 2026. By early 2027 expect $0.05–0.15/s for 1080p-with-audio via aggregators and 30-second single shots as standard. A daily wire-block-to-TikTok lane costs under a dollar a day — stamp the AIGC label automatically; C2PA detection catches it anyway.
Watch: confirmed Sora API death Sept 24 · rumored Kling 3.5/4.0 in Q4 · rumored Wan 3.x open weights eventually · speculative Veo 4 (may be absorbed into Omni entirely — don't plan around it).
The split is now clean: licensed rails versus litigation tails.
The flashy real-time systems are all platform-gated; the durable web layer is Gaussian splats.
The sweep's blind spot, caught in verification: Midjourney V8.1 (April — 5x faster, native 2K, better text), Flux 2 / Flux 2 Pro, Google Imagen 4, OpenAI GPT Image 2, and ByteDance's Seedream 4.5 with Seedream 5.0 previewed in February — multi-step visual reasoning with live web search during generation, a plausible fall GA. (Seedream is the image line; Seedance is video. The names will keep tripping people.)
R2 is confirmed-in-progress, held back over quality after a failed Huawei Ascend run. If it lands at DeepSeek pricing under V4's MIT license, reasoning reprices again like January 2025.
Prompt injection is operational in the wild at the exact moment x402 crosses 100M transactions and AP2 mandates roll out. One publicized theft via injected instructions freezes agentic-checkout adoption overnight and forces signed-mandate standards.
Hyperscaler capex is running $775–800B for 2026, far ahead of revenue. One guidance cut, power-buildout failure, or datacenter credit stress could snap it. Counter-signal: Microsoft's unfulfillable Azure backlog says demand still exceeds supply.
Evo generated complete novel phage genomes that outperform natural ones — published in Science. A misuse incident or pathogen-adjacent preprint triggers emergency screening mandates and open-weights restrictions for bio models.
Test-time-training research (context compressed into weights) plus agent-memory benchmarks are converging. A frontier lab shipping true per-user weight adaptation — an agent that genuinely learns your codebase — obsoletes the context-window race and much of RAG.
The Anthropic settlement dodged the outputs question; UMG's lyrics suit and Munich's ruling aim straight at generation. A US ruling that outputs infringe hits every agent that quotes or remixes web content — far broader than training liability.
X402_RECEIPT_SK and close the loop before the fall wave of paying agents. The receipt is a bet on the rail that landed under the Linux Foundation — but per the Instant Checkout autopsy, keep actual checkout on PointCast pages.