Blog

AI API relay provider selection, reviews, and industry news

Industry Analysis

Official Vendors Are Cutting Prices for Good — How Much Relay Margin Is Left?

How much relay margin is left now that official vendors are permanently cutting prices? This article answers it with verifiable official numbers: OpenAI went from GPT-6 Astra at $10/$50 to GPT-6 Sol at $2/$10 (declared permanent with no expiry date) to GPT-6 Luna at $0.10/$0.50 — a 100x output-price gap from Astra to Luna; GPT-6.1 Sol is also $2/$10, officially 'close to Astra at one-fifth the price,' with cached input at $0.10 (5% of list); Claude Opus 5.5 landed at $4/$20 with cache reads down 60%, and Sonnet 5.5 is cheaper still. We run the relay math: at the same 20% markup, gross spread per million output tokens falls from about ¥71 in the Astra era to about ¥14 at Sol pricing to about ¥0.7 at Luna pricing — while servers, bandwidth, payment rails, support and risk control don't get cheaper at all. Measured through the exchange-rate lever mainland relays commonly use (¥0.5/USD), the incentive for a user to register and prepay in order to save ¥3.3 evaporates at Luna pricing. Also covers the vendors' three downstream moves (OpenAI halving API credits per dollar on Pro and adding a $500/month Pro 500 tier; Alibaba's Token Plan covering Qwen/Kimi/GLM/DeepSeek in one subscription; Zhipu giving 100,000 users 100M tokens each), falsifies the 'industry is slowing down' narrative (at least 11 models shipped September 21-29, and lmmarketcap counts 60 September releases — yet GPT-6.1 Astra, set for October, was reported cancelled, and a DNS-tunneling incident paused tool-calling-related training, evaluation and inference for OpenAI's strongest models), and lays out what actually ships in October: Kimi K3.1 (official teaser page live, registry callable — strongly reported, media pointing at month-end, 1M context with Low/High/Max reasoning tiers), Qwen4 (officially 'coming soon,' no date), Gemini 4 (officially 'well before year-end' — no October claim), and Meta Watermelon (only an internal-document October window). Explicitly excluded: any DeepSeek V4.1 Pro schedule, the Space Bunny attribution, and any Grok 5 timeline. Lands on the five things relays must compete on: mainland direct connectivity and availability, aggregation the vendors structurally can't do, payment and invoicing, upstream channel quality and latency, and engineering that moves cost reduction from buying cheaper to using less (Nvidia's SoL-Pi at 50%/54.3% fewer tokens and $1,339→$894 per task; OpenSquilla retaining 99.96% quality at 88.9% lower cost; Anthropic cutting 80% of Claude Code's system prompt).

September 30, 2026 ~30 min read
Industry Analysis

Why Jev Went Viral: Inside the 1,954-Point Hacker News Surge — What's Real, and Which Numbers Are Marketing

This article doesn't re-explain what Jev is — it answers only why it spread so fast, and it separates heat from evidence. The trigger was TypeSafe's own September 15, 2026 announcement plus founder Diogo Almeida's launch thread (75,217 likes); an hour later the community pushed the post to 1,954 points and 511 comments on Hacker News (second for the week, first in AI). The site briefly lost the ability to serve users from its API because demand was so high (TechCrunch). Within 48 hours third parties put the launch video at 31M-40M views, and browser-use's jev-ultrafast hit 17,441 stars in a day. The hardest adoption signal is Vercel: Jev was the fastest-adopted model in AI Gateway history, reaching ~13% of teams on day one — 2x the GPT-5.6 family — and its engineer reports 5-18x faster and more accurate after switching off Luna. Add LangChain's Jev-as-a-Judge (0.44s average, $0.00035 per call) and elvex's 2,000 expense forms in 21 seconds for about 5 cents. But the heat exceeds the evidence: 193.6x / 444.6x are vendor self-reported upper bounds (TypeSafe disclosed three biases itself); Jev's accuracy is roughly 68% versus roughly 74% for GPT-5.6 Sol — it wins on cost, not accuracy; '0% hallucinations' is a schema guarantee, not factual correctness; independent tester Archestra found a 79% constant-baseline trap plus probability drift up to 0.17 run-to-run, and measured Sonnet 5 at 98% versus Jev at 93% on 100 real tool calls; the company admits it can't prove its pricing isn't subsidized; and the moat is thin — six clones appeared within two days and a 0.5B version runs on a MacBook. Chinese coverage lagged English by 3-6 days and was mostly compilation and tutorial content, with Tencent Tech Engineering's September 21 hands-on test the highest quality. A closing section offers a reusable ruler for judging the next viral model. The one thing mainland readers most need to know: Jev runs on POST /v1/systemone rather than an OpenAI-compatible endpoint, even OpenRouter needed a dedicated Decisions API channel, none of the 153 relays we catalog carries it, and the usual 'swap the base_url' trick fails at the root — leaving direct-from-vendor, OpenRouter, or a self-built gateway.

September 22, 2026 ~35 min read
Model Review

What Is "JEV"? TypeSafe AI's First System One Model, Jev — Shipped Features, Pricing, Release Timing, and How to Access It

Clearing up the ambiguity first: in an AI context, "JEV" means Jev, the first System One model from TypeSafe AI (founded by Diogo Almeida, an ex-OpenAI researcher who worked on ChatGPT's core training methods), announced September 15, 2026 — the other expansions of the acronym (Journal of Extracellular Vesicles, Japanese encephalitis vaccine) have nothing to do with AI. It generates no text at all, returning type-safe decisions: you send a state plus three kinds of structured questions (choice / score / noul) and get typed answers with calibrated probabilities and confidence. This article covers only what has actually shipped, per the official docs and pricing pages: the full spec of jev-1.13.0 (aliases jev-latest / jev-preview, both currently resolving to 1.13.0) — 64k context with 32k for state plus the longest question, text-only input (no image/audio/video), a 255 cardinality ceiling per field, no per-customer fine-tuning, requests not used for training, zero data retention available for enterprise, English as the primary training language with CJK working but unevenly; the exact response fields of each primitive (choice/probabilities/confidence, score/legend/probabilities/confidence, and noul on a 0-1 scale); RLCD (Reinforcement Learning for Calibrated Decisions) and the parallel sampler; pricing at $42 per billion input tokens ($0.042/M, output free, with TypeSafe claiming 238x lower input pricing than Claude Fable 5.1) and rate limits of 250,000 tokens/second and 1,200 requests/minute; the release timeline (official post September 15, homepage metadata September 17, OpenRouter listing September 18 at vendor-identical pricing); raw curl calls to POST https://api.typesafe.ai/v1/systemone plus complete pip install typesafe-sdk code, the JS SDK, and the official agent skill install commands. Section 9 is the key one: Jev is not an OpenAI-compatible chat.completions endpoint, so the usual 'swap the base_url' relay trick does not work — none of the 153 relay providers we catalog lists it, and the only third-party channel, OpenRouter, is not a direct-connect option for mainland China.

September 18, 2026 ~32 min read
Industry Analysis

Google's Gemini Lineup: Gemini 4 Pro Still Doesn't Exist — What Google Actually Sells (3.8 Flash, 3.8 Flash Cyber, 3.1 Pro, Omni 1.1 Flash), With Real Features, Real Pricing, Real Timing

Verified as of September 18, 2026: neither Gemini 4 nor Gemini 4 Pro appears in Google's official model catalog (page last updated 2026-09-17), and the changelog contains no fourth-generation entry of any kind. The only official statement is Pichai's July 26 remark about training 'much larger base models' — no date, no specs, no pricing; the sole source behind the circulating 'October launch' claim (Geeky Gadgets, September 9) misspells Anthropic as 'Enthropic' and calls Grok a SpaceX product, so we don't use it. This article covers what's real instead: Gemini 3.8 Flash (GA September 2 — DeepSWE v1.1 73.7%, Terminal-bench 2.1 89.4%, HLE-Verified 54.9%, Vals Finance Agent v2 61.4% ahead of Claude Opus 5, ~1M context / 64K output, $0.75/$3.75 promo doubling to $1.50/$7.50 in January 2027); the cybersecurity-focused 3.8 Flash Cyber, gated behind the Fairwind Program (CyberGym Pass@1 86.2%, 71.0% on a 20-language internal discovery benchmark); the 3.8 Live and 3.8 Live Extended Thinking voice models (GA September 15); the only purchasable Pro tier, 3.1 Pro (Preview, $2/$12 and up, HLE 44.4%, SWE-Bench Verified 80.6%, ARC-AGI-2 77.1%); the Omni 1.1 Flash video model (GA August 27 — extension, keyframe interpolation, 360p-4K); and the still-unreleased Gemini 3.5 Pro. Includes Google's itemized official pricing, the complete 2026 release timeline, and the mainland China relay situation across 153 cataloged providers — ProAI API explicitly lists gemini-3.7-flash as live, MKEAI/NodAPI/Quanzil cover Gemini 3.5-3.7, 14 carry gemini-3.1-pro, but none lists gemini-3.8-flash — with OpenAI-compatible and LiteLLM access code.

September 18, 2026 ~26 min read
Industry Analysis

GPT-6 Sol Preview: A 6x-Faster 'Speed Tier' That Would Complete the GPT-6 Family

A forward look at OpenAI's GPT-6 Sol as of September 2026: per a QbitAI report dated 2026-09-07, OpenAI is internally testing a speed tier positioned below Astra — about 3 minutes for Sol versus roughly 19 for Astra on one SVG-generation task (about 6x), quality slightly below Astra's but still 'monster-level,' possibly launching at the September 29 developer conference. Why the GPT-6 family is missing exactly this tier (Astra shipped September 4 at $10/$50 with ARC-AGI-3 99.9%, GPQA Diamond 96%, OSWorld 2.0 72.6%, and new $200/month Pro sign-ups paused around September 10); what a 6x speed advantage buys inside agent loops, batch generation, and draft rounds; where Sol would likely price between Astra and the volume tiers like Gemini 3.7 Flash (projection, not fact); how it would reach you through official API → OpenRouter / AiHubMix → mainland China relays and which providers list first; the three checks to run in its first week (model ID reconciliation, your own workload baseline, price sanity); the September 29 Dev Day watch list and three scenarios; and five things to prepare now — with OpenAI-compatible access code and a LiteLLM fast-tier/flagship routing config.

September 15, 2026 ~15 min read
Industry Analysis

Meta's Watermelon Model Explained: The Next Llama, or a Closed Flagship Behind HATCH?

A breakdown of Watermelon, Meta's next flagship AI model, based on the August 25, 2026 BlockBeats report citing The Information and Meta internal documents: what it is (an internal codename widely reported for an October 2026 release), its relationship to the HATCH / OpenClaw consumer agent platform (an Instagram shopping assistant with a premium tier up to $199.99/month), the two possible routes for open weights vs. closed (a Llama-style open successor vs. a Muse Spark-style closed flagship), why 'no public specs yet' is itself meaningful, and how to expect to reach it through AI API relays and AI gateways — self-hosting (Ollama/vLLM), aggregators (Groq/Together/SiliconFlow), mainland China relays, or the Meta Model API (api.meta.ai/v1) + OpenRouter if closed. Every claim tagged: reported / inference / unverified.

August 28, 2026 ~13 min read
Industry Analysis

DeepSeek V4-Flash-Vision-Exp Explained: The Multimodal Flash Posts Its First Agent Benchmarks, Drawing Opus 4.8 2:2 — What the Vision Variant Means and How to Reach It

A breakdown of V4-Flash-Vision-Exp, the vision-enabled experimental variant of DeepSeek V4 Flash announced August 22, 2026: multimodal input added to the 284B/13B MoE, 1M-context base; the first official Agent benchmarks that tie Anthropic's Opus 4.8 2:2 across four multimodal evals (wins on Agents' Last Exam 27.3 vs 25.7 and ZeroBench 35.0 vs 34.0; losses on ApexBench 36.5 vs 39.4 and Chartography 64.3 vs 65.0); the jump over the text-only V4-Flash-0731 (ApexBench 26.2→36.5, DeepSWE 54.4→59.3, ahead of Opus 4.8's 58.0); why pricing most likely mirrors V4 Flash's peak/off-peak tier (off-peak ¥1.5/¥4.5, peak ¥3/¥9 per M tokens) though separate vision-input pricing is unverified; and the mainland China relay situation (none of the 153 cataloged providers lists the vision model ID yet; fast-syncing relays like SiliconFlow, ProAI API, NoDAPI and JENIYA are the likely first movers) with OpenAI-compatible access code.

August 22, 2026 ~21 min read
Industry Analysis

Google Gemini 3.7 Flash Review: The Coding-and-Agent 'Workhorse' That Iterates Every Three Weeks — How to Read (and Reach) the $0.75/M Promo

A deep dive into Gemini 3.7 Flash, released August 14, 2026 — billed as Google's 'most intelligent workhorse model yet,' built for coding and agentic workflows and arriving just three weeks after Gemini 3.6 Flash: official specs (~1M input context, 64K output, native multimodal input), the gap between Google-reported benchmarks (DeepSWE v1.1 49.0%→65.3%, FrontierCode 1.1 34.4%→43.6%, narrowly ahead of Claude Sonnet 5) and the still-missing third-party AAII score, the $0.75/$3.75 six-month promo (back to $1.50/$7.50 in January) that is actually shared across the whole Flash line, honest per-token comparison with Claude Sonnet 5, GPT-5.6 Terra, DeepSeek V4 Flash/Pro, Kimi K3, GLM-5.3 and Qwen3.8-Max, and how to reach it through mainland China relay providers (ProAI API explicitly live, MKEAI/NoDAPI/JENIYA covering Gemini 3.7) with OpenAI-compatible code.

August 19, 2026 ~20 min read
Industry Analysis

GLM-5.3 Review: The Strongest Open-Source Coding Model? A Flagship Upgrade Built Entirely on Post-Training

A deep dive into Zhipu's GLM-5.3, released August 14, 2026: 743B parameters, the same base as GLM-5.2, unchanged architecture — yet coding and agent ability jumped dramatically (DeepSWE v1.1 46.2→66.9, #1 open-source; Terminal-Bench 3.0 4.6→28.3; CyberGym 84.5%, #1 among all evaluated models, ahead of Mythos 5's 83.8% and GPT-5.6 Sol's 83.6%). An honest comparison of the real gap versus closed flagships Fable 5 / GPT-5.6 Sol and GLM-5.3's token-efficiency edge, the double-edged sword of its cybersecurity capability ('trusted access' gating, API requiring thinking enabled), and how to use it via a relay provider and when (GLM Coding Plan / ZCode / AutoClaw available now; API + weights in ~2 weeks).

August 15, 2026 ~20 min read
Industry Analysis

DeepSeek Harness Architecture Deep Dive: What Cordis's 'Revertible Effects' and 'Reactive Coeffects' Actually Mean, and the Mechanical Details of the Four Run Modes

One layer below 'everything is a plugin,' this piece dissects Cordis, the plugin framework underneath DeepSeek Harness: its lineage (the shigma/Koishi ecosystem), the two key concepts from the paper A Programming Paradigm for Spatiotemporal Composability — Revertible Effects (temporal) and Reactive Coeffects (spatial) — and the mechanics of effect/disposer/fiber; then how DSH turns the model adapter, tool registry, session log, and the agent loop itself into swappable plugins, and finally the mechanical details of the four run modes (standard, PTC/Code, minimal, cordis/Creation).

August 14, 2026 ~22 min read
Industry Analysis

DeepSeek Harness Deep Dive: DeepSeek's Agent Is Here — But It Released an 'Agent-Assembly Runtime,' Not Another Codex

A breakdown of DeepSeek Harness (DSH), the agent product DeepSeek opened to developers on August 13, 2026 as an MIT-licensed developer preview v0.1: what it actually is (an open-source agent runtime, not a closed coding agent), what it does (a local agent workbench, four run modes, an MCP client bridge, append-only session logs), what framework it's built on (the Cordis plugin system — 'everything is a plugin'), what's genuinely new versus merely re-wrapped, and — most practically — how to point it at a mainland China AI API relay via DEEPSEEK_BASE_URL, the Web UI, or settings.yaml so DeepSeek V4 Pro / V4 Flash do the work.

August 14, 2026 ~19 min read
Industry Analysis

DeepSeek V4 Pro (GA) vs Grok 4.6: Same Night, Two 'Half-Price Flagships' — Which One Should You Integrate?

A head-to-head of the two flagships that landed the same night — DeepSeek V4 Pro GA (1.6T/49B MoE, 1M context, dual OpenAI/Anthropic compatibility) and Grok 4.6 (≈1.5T params, AAII 61 matching GPT-5.6 Sol Max): real specs, the gap between vendor-reported benchmarks and third-party AAII (53 vs 61), the pricing collision (≈7x output gap, 100x+ cache-hit gap), the ecosystem battle, and how both are actually reachable through mainland China AI API relay providers.

August 13, 2026 ~20 min read
Tutorial

DeepSeek V4 Flash Free-Access Quick-Start Guide: Squeeze Every Free Token

Which platforms actually serve deepseek-v4-flash and how to connect fast and use their free tokens: SenseNova Token Plan free beta, OpenCode Zen limited-time free, NVIDIA NIM free prototyping, DeepSeek official's 5M-token signup bonus, plus free quotas from Alibaba Bailian, Tencent TokenHub, Qiniu Cloud and ModelScope — and 10 second-tier providers. Includes a base_url + model-ID cheat sheet and an OpenAI-compatible code snippet, with an honest look at the hard limits on every free path.

August 12, 2026 ~20 min read
Industry Analysis

Meta Muse Spark 1.2 Review: The Coding-Agent Contender That Wins on Cost-Efficiency

A deep dive into Meta Superintelligence Labs' (MSL) August 5, 2026 release of Muse Spark 1.2 and its terminal coding agent Muse Code: what kind of model it actually is, the cost-per-intelligence numbers, the persistent background-agent architecture, the 1M-token multimodal context window, pricing, and how big the gap is between Meta's self-reported benchmarks and independent verification from Vals AI and others.

August 8, 2026 ~16 min read
Industry Analysis

The AI Video Model Pricing Comparison: Seedance 2.0/2.5, Kling 3.0, Veo 3.1, and Wan 2.7 — Which One Is Actually Cheapest

A head-to-head breakdown of pricing across four leading video generation models: ByteDance's Seedance 2.0 and 2.5 (whose API pricing dropped the day this piece went to print), Kuaishou's Kling 3.0, Google's official three-tier Veo 3.1 rates, and Alibaba's open-weight Wan 2.7. Covers both access routes, a real 20%+ price spread for the same Seedance 2.0 spec across platforms, which of eggstriker's reviewed vendors actually sell these four models, and why only one of the four genuinely can't be reached without a proxy from mainland China.

August 7, 2026 ~19 min read
Industry Analysis

DeepSeek Announced a Price Hike Without Saying How Much: What We Verified, How the Community Is Reacting, and Whether Relay 'Savings' Are Real

On August 6, DeepSeek posted a vague, number-free notice announcing a broad price increase, while July's rumored peak-hour surge pricing still isn't live as of publication. We re-checked the official pricing page in real time, surveyed reactions on V2EX, Zhihu, and Reddit, broke down the technical mitigations developers are actually using, and cross-checked what SiliconFlow, EasyRouter, and TokenRiver's 'cheaper than official' claims actually mean — three completely different mechanisms, only some of which genuinely save money on DeepSeek specifically.

August 7, 2026 ~12 min read
Industry Analysis

The Complete AI Model Pricing Comparison: OpenAI, Anthropic, Google, DeepSeek, Kimi, Qwen, GLM, xAI, and MiniMax, Side by Side

A head-to-head breakdown of current flagship pricing across nine major AI labs: per-million-token input/output rates, context windows, caching discounts, DeepSeek's new peak-hour surge pricing, how forced 'always-on thinking' quietly inflates bills, and why open-weight and closed models follow completely different pricing logic once relay providers get involved. Includes Claude Opus 5 quietly replacing Opus 4.8 at the same price, Qwen3.8-Max's just-announced GA pricing, and Gemini 3.5 Pro's still-unreleased status — independently re-verified before publication.

August 5, 2026 ~20 min read
Industry Analysis

OpenAI Teases "Astra": The Next Model Family, Announced Almost In Passing, Inside a Math Paper

A breakdown of OpenAI's next-generation model Astra: how researcher Noam Brown revealed it on August 1, 2026 buried inside a blog post about math results, the 10 long-standing math and theoretical-CS problems it helped solve, Sam Altman's demo to Washington policymakers, the incoming Trump-administration pre-release review framework, Gary Marcus's skepticism, and what none of this means yet for API pricing or mainland China relay access.

August 3, 2026 ~16 min read
Industry Analysis

MiniMax H3 Open Weights vs. ByteDance's Closed Seedance 2.5: A Same-Day Launch, Two Opposite Bets

On July 31, 2026, MiniMax's H3 and ByteDance's Seedance 2.5 launched on the same day, but staked out opposite strategies: one promising open weights, the other staying closed-API only. A breakdown of H3's real architecture and specs, how much of MiniMax's 'open-weight' promise has actually held up over the past year, why ByteDance keeps betting closed, where this fits in China's broader open-vs-closed AI landscape, plus H3's real rankings on Artificial Analysis's three video leaderboards, its pricing, and how it's actually accessible through mainland China AI API relay providers.

August 1, 2026 ~19 min read
Industry Analysis

Seedance 2.5 Deep-Dive Review: Single-Shot Length Doubles to 30 Seconds, But Official API Pricing Is Still a Blank

A breakdown of ByteDance's real Seedance 2.5 upgrades (single-shot length doubling from 15s to 30s, reference inputs expanding to 50, four new editing modes), the full timeline from June teasers to the July 31 official launch, how it stacks up against Kling 3.0/Veo 3.1/a fading Sora 2, the still-unannounced official API pricing versus conflicting third-party estimates, the unresolved Hollywood copyright dispute, and how it's actually accessible through mainland China AI API relay providers.

August 1, 2026 ~15 min read
Industry Analysis

Portkey Deep-Dive Review: Gateway + Observability + Guardrails in One — Still Worth It After the Palo Alto Networks Acquisition?

A breakdown of Portkey's core features (AI Gateway, Virtual Keys, observability, 60+ guardrails, governance, prompt management, MCP Gateway), the full timeline from its March 2026 gateway open-sourcing to its ~$700M acquisition by Palo Alto Networks into Prisma AIRS, real pricing (Developer free / Production $49/mo / Enterprise custom), and how it actually relates to LiteLLM, OpenRouter, and mainland China AI API relay providers.

July 28, 2026 ~22 min read
Tutorial

The Complete Jina API Guide (2026): Why RAG Developers Still Care About It — A Competitive Breakdown and a Hands-On Domain Knowledge Graph Tutorial

Jina AI is the retrieval-layer bundle that kept running independently after being acquired by Elastic — Embeddings, Reranker, and Reader all share the same token pool. An honest breakdown of its strengths and weaknesses versus Firecrawl/Tavily/Exa/Diffbot, plus a tested pipeline that shows you how to actually turn web pages into a queryable domain knowledge graph, with full Python code.

July 19, 2026 ~24 min read
Industry Analysis

Is Grok 4.5 Really the Best-Value AI Model Right Now? Ranking the Most Cost-Effective Models in July 2026

Grok 4.5 ($2/M input, $6/M output) nearly matches Claude Opus 4.8 on benchmarks, uses roughly 4x fewer output tokens for the same work, and leads every model on agentic tool use — but it's neither the single cheapest model nor the cheapest multimodal model. We rank 8 mainstream models on real pricing and honestly break down where each one actually stands, plus mainland China relay access for each.

July 16, 2026 ~17 min read