What changed, and when
Follow newly published and updated guides, site releases, model and pricing updates, and edits to tracked provider values. Model updates appear when approved data is deployed; pending imports are not published changes.
New and updated guides
Published guide releases, newest first. Drafts and scheduled guides appear only after publication.
-
Decide whether your agents need an MCP gateway. Compare tool discovery, authorization, upstream credentials and audit trails with a practical acceptance checklist.
-
Check coding-agent gateway compatibility beyond a base URL: API formats, streaming, tools, model aliases, context, credentials and team rollout controls.
-
Plan a model migration with a version inventory, contract tests, quality gates, canary traffic and an executable rollback. Distinguish aliases from real model identity.
-
Learn when LLM routing reduces prompt cache hits, how sticky routing helps, and how to compare cached-token costs with a repeatable routing test.
-
Compare LLM gateway spending limits: what they block, whether BYOK counts, how concurrent requests overshoot, and how to test the cutoff before an agent runs unattended.
Site releases and model updates
Other publication items from the latest 300 recorded updates. Model comparisons start with the September 16, 2026 release. Counts describe changes to published listings, not necessarily vendor launch dates.
-
A new sourced head-to-head comparison is available.
-
A new sourced head-to-head comparison is available.
-
A new sourced head-to-head comparison is available.
-
A new sourced head-to-head comparison is available.
-
A new sourced head-to-head comparison is available.
-
Filter models across providers by documented Chinese-developed families, including Qwen, Z.ai GLM, DeepSeek, Moonshot Kimi, MiniMax, Tencent Hunyuan, Baidu ERNIE, Yi, InternLM, BGE and Wan. Named derivatives are included; developer origin does not specify hosting location or data residency.
-
What changed now includes site release notes and summaries of model additions, removals and pricing changes when approved data is published. These updates also become available for subscriber digests. Model change tracking starts with this release; older model changes are not reconstructed. Email editions still require approval before sending.
-
Search models across providers, browse each provider’s model catalog, and compare published input/output token rates. Sort by newest or popularity and filter by use, free models, and OpenRouter-sourced collections. Prices retain route, region and upstream-reference qualifications. This note documents features already available on the site; it does not announce pending model imports.
Provider field edit history
-
9 changes
- New to you 120 94
Live endpoint returned 94 models (was 120, a change of -26). Counted directly from the vendor's authenticated account catalog API.
Source for this change ↗ - New to you 274 3
Live endpoint returned 3 models (was 274, a change of -271). Counted directly from the vendor's authenticated account catalog API.
Source for this change ↗ - New to you 9273 9458
GitHub reports 9,458 stars (was 9,273).
Source for this change ↗ - New to you 200 272
Live endpoint returned 272 models (was 200, a change of +72). Counted directly from the vendor's authenticated account catalog API.
Source for this change ↗ - New to you Not published 11
Live endpoint returned 11 models (was 0, a change of +11). Counted directly from the vendor's authenticated account catalog API.
Source for this change ↗ - New to you 100 27
Live endpoint returned 27 models (was 100, a change of -73). Counted directly from the vendor's authenticated account catalog API.
Source for this change ↗ - New to you 1066 1104
Live endpoint returned 1104 models (was 1066, a change of +38). Counted directly from the vendor's own models API.
Source for this change ↗ - New to you 445 458
Live endpoint returned 458 models (was 445, a change of +13). Counted directly from the vendor's own models API.
Source for this change ↗ - New to you 8142 8359
GitHub reports 8,359 stars (was 8,142).
Source for this change ↗
- New to you
-
5 changes
- New to you 143 136
Live endpoint returned 136 models (was 143, a change of -7). Counted directly from the vendor's own models API.
Source for this change ↗ - New to you 373 386
Live endpoint returned 386 models (was 373, a change of +13). Counted directly from the vendor's own models API.
Source for this change ↗ - New to you 16800 17155
GitHub reports 17,155 stars (was 16,800).
Source for this change ↗ - New to you 4845 4985
GitHub reports 4,985 stars (was 4,845).
Source for this change ↗ - New to you 1282 1525
Live endpoint returned 1525 models (was 1282, a change of +243). Counted directly from the vendor's own models API.
Source for this change ↗
- New to you
-
1 change
- New to you Not published Managed marketplace
Owner-requested onboarding; retention conflict and unknown fees preserved. Static publication requires deployment.
Source for this change ↗
- New to you
-
12 changes
- New to you 1027 1066
Live endpoint returned 1066 models (was 1027, a change of +39). Counted directly from the vendor's own models API.
Source for this change ↗ - New to you 136 143
Live endpoint returned 143 models (was 136, a change of +7). Counted directly from the vendor's own models API.
Source for this change ↗ - New to you 429 445
Live endpoint returned 445 models (was 429, a change of +16). Counted directly from the vendor's own models API.
Source for this change ↗ - New to you Not published 1282
Live endpoint returned 1282 models (was 0, a change of +1282). Counted directly from the vendor's own models API.
Source for this change ↗ - New to you 7600 8142
GitHub reports 8,142 stars (was 7,600).
Source for this change ↗ - New to you 601 1643
GitHub reports 1,643 stars (was 601).
Source for this change ↗ - New to you 57500 59000
GitHub reports 59,000 stars (was 57,500).
Source for this change ↗ - New to you 1987 2112
GitHub reports 2,112 stars (was 1,987).
Source for this change ↗ - New to you 47090 48314
GitHub reports 48,314 stars (was 47,090).
Source for this change ↗ - New to you Not published 211
GitHub reports 211 stars (was 0).
Source for this change ↗ - New to you Proprietary Apache-2.0
GitHub reports the licence as Apache-2.0, but the catalog says Proprietary. A relicence is a material change — verify against the repository before accepting.
Source for this change ↗ - New to you Not published AI Gateway HQ
Owner-requested addition. Managed gateway; sourced research dated 17 September 2026. Private deployment not GA; no certification or SLA assumed.
Source for this change ↗
- New to you
-
7 changes
- New to you not_documented otel
Local snapshot import; verify against the linked source before accepting.
Source for this change ↗ - New to you `n.a.` — No OpenTelemetry or agent tracing documented for the gateway ([Observability](https://vercel.com/docs/ai-gateway/observability-and-spend/observability)) Native OTLP/HTTP Trace Drains on Pro and Enterprise; separate delivery and egress charges. Request, routing, model-attempt and provider spans contain metadata, not prompt or completion content. Request traces do not establish complete application or agent tracing.
Local snapshot import; verify against the linked source before accepting.
Source for this change ↗ - New to you 20 0
Local snapshot import; verify against the linked source before accepting.
Source for this change ↗ - New to you Credit-based pay-as-you-go with 0% markup and no platform fee on tokens. Monetizes gateway features à la carte: provider allowlists, ZDR, custom reporting writes/queries, Trace Drains. Enterprise can pay by invoice with no processing fees. Pricing page last updated 2026-08-23. Gateway credits are pay-as-you-go with no token markup or mandatory platform subscription. Optional governance, reporting and trace exports have separate charges; some require Pro or Enterprise. Include only incremental platform costs for the selected features.
Local snapshot import; verify against the linked source before accepting.
Source for this change ↗ - New to you [{"label":"Team-wide zero data retention","amount":"$0.10 per 1,000 requests (Pro and Enterprise)"},{"label":"Team-wide provider allowlist","amount":"$0.10 per 1,000 successful requests (Pro and Enterprise)"},{"label":"Custom Reporting writes","amount":"$0.075 per 1,000 tag / user ID / quota entity writes"},{"label":"Custom Reporting queries","amount":"$5 per 1,000 queries to the reporting endpoint"},{"label":"Vercel Pro developer seat","amount":"$20 per month"}] [{"label":"Team-wide zero data retention","amount":"$0.10 per 1,000 requests (Pro and Enterprise)"},{"label":"Team-wide provider allowlist","amount":"$0.10 per 1,000 successful requests (Pro and Enterprise)"},{"label":"Custom Reporting writes","amount":"$0.075 per 1,000 tag / user ID / quota entity writes"},{"label":"Custom Reporting queries","amount":"$5 per 1,000 queries to the reporting endpoint"},{"label":"Optional Vercel Pro developer seat","amount":"$20 per month"},{"label":"Optional Trace Drains delivery","amount":"$0.05 per 1,000 traces (Pro and Enterprise)"},{"label":"Optional Trace Drains egress","amount":"$0.50 per GB (Pro and Enterprise)"}]
Local snapshot import; verify against the linked source before accepting.
Source for this change ↗ - New to you not_documented otel
Local snapshot import; verify against the linked source before accepting.
Source for this change ↗ - New to you `n.a.` — no OpenTelemetry or agent-trace representation documented for AI Gateway ([Logging](https://developers.cloudflare.com/ai-gateway/observability/logging/)) Configurable native OTLP export in JSON or Protobuf. Documented attributes include model, provider, usage, cost, custom metadata and prompt/completion payloads. Review content handling and collector access before enabling exports; validate parent-context propagation in your application.
Local snapshot import; verify against the linked source before accepting.
Source for this change ↗
- New to you
-
9 changes
- New to you 500 429
Live endpoint returned 429 models (was 500, a change of -71). Counted directly from the vendor's own models API.
Source for this change ↗ - New to you 200 136
Live endpoint returned 136 models (was 200, a change of -64). Counted directly from the vendor's own models API.
Source for this change ↗ - New to you 600 706
Live endpoint returned 706 models (was 600, a change of +106). Counted directly from the vendor's own models API.
Source for this change ↗ - New to you 500 1027
Live endpoint returned 1027 models (was 500, a change of +527). Counted directly from the vendor's own models API.
Source for this change ↗ - New to you 360 373
Live endpoint returned 373 models (was 360, a change of +13). Counted directly from the vendor's own models API.
Source for this change ↗ - New to you Open core Apache-2.0
GitHub reports the licence as Apache-2.0, but the catalog says Open core. A relicence is a material change — verify against the repository before accepting.
Source for this change ↗ - New to you Open core MIT
GitHub reports the licence as MIT, but the catalog says Open core. A relicence is a material change — verify against the repository before accepting.
Source for this change ↗ - New to you 4691 4845
GitHub reports 4,845 stars (was 4,691).
Source for this change ↗ - New to you Open core MIT
GitHub reports the licence as MIT, but the catalog says Open core. A relicence is a material change — verify against the repository before accepting.
Source for this change ↗
- New to you
-
17 changes
- New to you Inputs and outputs are not stored by default. Temporary caching may be used to improve performance unless configured otherwise. Prompts and responses are stored by default for product improvements. Organization administrators can disable storage to enable ZDR; training is a separate opt-in.
Owner-authorized high-priority audit correction; source reviewed 2026-09-05.
Source for this change ↗ - New to you Yes Not published
Owner-authorized high-priority audit correction; source reviewed 2026-09-05.
Source for this change ↗ - New to you Not published Yes
Owner-authorized high-priority audit correction; source reviewed 2026-09-05.
Source for this change ↗ - New to you Zero data retention is the default posture, not an upgrade. Zero retention is the default for open-model inference without opt-in logging. Responses API storage is on by default for 30 days; set store=False to disable it.
Owner-authorized high-priority audit correction; source reviewed 2026-09-05.
Source for this change ↗ - New to you yes depends
Owner-authorized high-priority audit correction; source reviewed 2026-09-05.
Source for this change ↗ - New to you On by default and user-controlled, but note the boundary: ZDR applies only from the moment you enable it and does nothing about data already processed. ZDR must be enabled by disabling prompt/response storage in organization privacy settings. It applies prospectively and disables passthrough models.
Owner-authorized high-priority audit correction; source reviewed 2026-09-05.
Source for this change ↗ - New to you none full_content
Owner-authorized high-priority audit correction; source reviewed 2026-09-05.
Source for this change ↗ - New to you 0 Not published
Owner-authorized high-priority audit correction; source reviewed 2026-09-05.
Source for this change ↗ - New to you Zero by default. No retention window is published for the opt-in storage path. Prompt storage is on by default; a retention duration is not published on the privacy page. Enabling organization ZDR stops persistence of future request content.
Owner-authorized high-priority audit correction; source reviewed 2026-09-05.
Source for this change ↗ - New to you No prompt retention by default; endpoint metrics/events retention is not stated ([Privacy and security](https://docs.together.ai/docs/privacy-and-security), [Monitor endpoints and deployments](https://docs.together.ai/docs/dedicated-endpoints/monitoring)) Default prompt storage has no published duration on the reviewed privacy page. Organization ZDR disables future content storage; endpoint event retention is not specified. [Privacy and security](https://docs.together.ai/docs/privacy-and-security).
Owner-authorized high-priority audit correction; source reviewed 2026-09-05.
Source for this change ↗ - New to you 90 Not published
Owner-authorized high-priority audit correction; source reviewed 2026-09-05.
Source for this change ↗ - New to you Ninety days for logs and 365 for metrics by default. Self-hosted puts logs in your own S3-compatible store under your lifecycle policy. Plan and deployment dependent: Developer logs 3 days; Production logs 30 days ([Logs](https://portkey.ai/docs/product/observability/logs)). SaaS Enterprise logs 90 days and metrics 365 days; Hybrid logs follow customer storage lifecycle rules ([Security](https://portkey.ai/docs/enterprise/security)). No single default applies to all plans.
Owner-authorized high-priority audit correction; source reviewed 2026-09-05.
Source for this change ↗ - New to you Developer 3 days (10k logs/month), Production 30 days (100k logs/month, $9 per additional 100k), Enterprise unlimited logs with retention unspecified ([Logs](https://docs.portkey.ai/docs/product/observability/logs)) Plan and deployment dependent: Developer logs 3 days; Production logs 30 days ([Logs](https://portkey.ai/docs/product/observability/logs)). SaaS Enterprise logs 90 days and metrics 365 days; Hybrid logs follow customer storage lifecycle rules ([Security](https://portkey.ai/docs/enterprise/security)). No single default applies to all plans.
Owner-authorized high-priority audit correction; source reviewed 2026-09-05.
Source for this change ↗ - New to you otel not_documented
Owner-authorized high-priority audit correction; source reviewed 2026-09-05.
Source for this change ↗ - New to you otel not_documented
Owner-authorized high-priority audit correction; source reviewed 2026-09-05.
Source for this change ↗ - New to you otel otel_partial
Owner-authorized high-priority audit correction; source reviewed 2026-09-05.
Source for this change ↗ - New to you otel otel_integration
Owner-authorized high-priority audit correction; source reviewed 2026-09-05.
Source for this change ↗
- New to you
-
1 change
- New to you
New catalog entry: zero-markup routing layer from Hugging Face (proprietary service, base URL router.huggingface.co/v1). Fans a single HF token out to 18 third-party inference partners, 15 of which serve chat completion, with automatic failover and :fastest / :cheapest / :preferred routing policies. Hugging Face charges the same rates as the provider with no additional fees. Scoped to the Inference Providers product only: the Hub and Inference Endpoints are out of scope, and Hub-level SOC 2, GDPR and EU-residency statements are deliberately not carried over to the gateway fields.
Source for this change ↗
- New to you
-
7 changes
- New to you
New catalog entry: Self-host-only LLM key-management gateway (AGPL-3.0, hard-fork derivative of One API, 47,090 GitHub stars, v1.0.0-rc.30 released 2026-08-31). Highest star count of any candidate. Two GitHub advisories on record: CVE-2026-71479 (quota-overflow self-crediting) and CVE-2026-64859 (root token leak), both patched.
Source for this change ↗ - New to you
New catalog entry: LLM gateway feature inside MLflow Tracking Server (Apache-2.0, Linux Foundation, MLflow 3.15.2 released 2026-08-26). Correcting the deprecation myth: renamed to Deployments Server in 2.9.0, then explicitly reverted in 2.17.0 (2024-10-11); MLflow 3.0 removed the standalone deployment-server app but the gateway continues inside Tracking Server.
Source for this change ↗ - New to you
New catalog entry: Managed multi-modality aggregator (Proprietary, 500+ models across 50+ providers, `POST /v3/chat/completions` at api.edenai.run, 5.5% platform fee at checkout with no markup on provider pricing). Not LLM-routing-first: half the product is non-LLM expert models (OCR, speech, document parsing).
Source for this change ↗ - New to you
New catalog entry: Linux Foundation Rust-based data plane routing HTTP, gRPC, LLM, MCP and A2A traffic (Apache-2.0, self-host, drop-in OpenAI-compatible, v1.5.0 released 2026-08-27). Solo Enterprise for agentgateway is the same project as a commercial distribution.
Source for this change ↗ - New to you
New catalog entry: CNCF Sandbox AI-native API gateway on Istio/Envoy with Wasm plugins (Apache-2.0, self-host + commercial Alibaba Cloud edition, 31 providers via AI Proxy plugin, v2.2.4 released 2026-08-13).
Source for this change ↗ - New to you
New catalog entry: Managed LLM gateway from Merge (Proprietary, launched 31 March 2026, base URL api-gateway.merge.dev/v1). Routes to 24 model providers across 274 models, priced as LLM cost plus a 5% fee on the Pro plan. Distinct product line from Merge Unified; compliance material scoped to Unified is not being carried over.
Source for this change ↗ - New to you
New catalog entry: Kubernetes-native OSS AI gateway from the Envoy Gateway project (Apache-2.0, self-host, Helm/K8s, `POST /v1/chat/completions` fully supported, v1.1.0 released 2026-08-21).
Source for this change ↗
- New to you
-
3 changes
- New to you 1000 Not published
Correction to our own reading, not a vendor change. We could not trace 1,000 to any page Bifrost publishes. The overview states "20+ AI providers" and names example models rather than a catalogue, so provider breadth is the only figure the vendor stands behind.
Source for this change ↗ - New to you 1892 Not published
Correction to our own reading, not a vendor change. We could not trace 1,892 to any page LiteLLM publishes. The providers documentation states no aggregate total and expresses coverage per provider; the repository says "100+ LLMs" and, confusingly, "100+ LLM providers" for the same figure. The field is now blank, which on this site reads "not published" rather than "no", and provider breadth is used for comparisons instead.
Source for this change ↗ - New to you 19 110-200
Correction to our own reading, not a vendor change. The earlier figure of 19 was counted from the example models named on the docs overview page, which is an illustrative list rather than the catalogue. Together publishes "200+ models" in its model directory and enumerates 110 on serverless; the range now carries both, because serverless is a subset of what the platform will run.
Source for this change ↗
- New to you
-
20 changes
- New to you
Initial research pass. Date is the earliest per-field verification recorded for this product; the time of day is an ordering key, not evidence.
Source for this change ↗ - New to you
Initial research pass. Date is the earliest per-field verification recorded for this product; the time of day is an ordering key, not evidence.
Source for this change ↗ - New to you
Initial research pass. Date is the earliest per-field verification recorded for this product; the time of day is an ordering key, not evidence.
Source for this change ↗ - New to you
Initial research pass. Date is the earliest per-field verification recorded for this product; the time of day is an ordering key, not evidence.
Source for this change ↗ - New to you
Initial research pass. Date is the earliest per-field verification recorded for this product; the time of day is an ordering key, not evidence.
Source for this change ↗ - New to you
Initial research pass. Date is the earliest per-field verification recorded for this product; the time of day is an ordering key, not evidence.
Source for this change ↗ - New to you
Initial research pass. Date is the earliest per-field verification recorded for this product; the time of day is an ordering key, not evidence.
Source for this change ↗ - New to you
Initial research pass. Date is the earliest per-field verification recorded for this product; the time of day is an ordering key, not evidence.
Source for this change ↗ - New to you
Initial research pass. Date is the earliest per-field verification recorded for this product; the time of day is an ordering key, not evidence.
Source for this change ↗ - New to you
Initial research pass. Date is the earliest per-field verification recorded for this product; the time of day is an ordering key, not evidence.
Source for this change ↗ - New to you
Initial research pass. Date is the earliest per-field verification recorded for this product; the time of day is an ordering key, not evidence.
Source for this change ↗ - New to you
Initial research pass. Date is the earliest per-field verification recorded for this product; the time of day is an ordering key, not evidence.
Source for this change ↗ - New to you
Initial research pass. Date is the earliest per-field verification recorded for this product; the time of day is an ordering key, not evidence.
Source for this change ↗ - New to you
Initial research pass. Date is the earliest per-field verification recorded for this product; the time of day is an ordering key, not evidence.
Source for this change ↗ - New to you
Initial research pass. Date is the earliest per-field verification recorded for this product; the time of day is an ordering key, not evidence.
Source for this change ↗ - New to you
Initial research pass. Date is the earliest per-field verification recorded for this product; the time of day is an ordering key, not evidence.
Source for this change ↗ - New to you
Initial research pass. Date is the earliest per-field verification recorded for this product; the time of day is an ordering key, not evidence.
Source for this change ↗ - New to you
Initial research pass. Date is the earliest per-field verification recorded for this product; the time of day is an ordering key, not evidence.
Source for this change ↗ - New to you
Initial research pass. Date is the earliest per-field verification recorded for this product; the time of day is an ordering key, not evidence.
Source for this change ↗ - New to you
Initial research pass. Date is the earliest per-field verification recorded for this product; the time of day is an ordering key, not evidence.
Source for this change ↗
- New to you
When each product was last verified
Each product's date is its oldest field check, not its newest — a single stale figure is what will mislead you, so that is what gets reported. Sourced fields counts how many values carry a citation: a low number means thin coverage, not that the product lacks the feature.
| Product | Type | Sourced fields | Oldest check | Age |
|---|---|---|---|---|
| Apache APISIX AI Gateway | Open source | 29 | 27d | |
| Google Vertex AI | Cloud platform | 32 | 27d | |
| Kong AI Gateway | Managed gateway | 34 | 27d | |
| LLM Gateway | Open source | 34 | 27d | |
| Groq | Inference provider | 34 | 27d | |
| Amazon Bedrock | Cloud platform | 35 | 27d | |
| Azure AI Foundry | Cloud platform | 35 | 27d | |
| Cloudflare AI Gateway | Cloud platform | 36 | 27d | |
| Fireworks AI | Inference provider | 36 | 27d | |
| TrueFoundry AI Gateway | Cloud platform | 37 | 27d | |
| Braintrust Gateway | Managed gateway | 39 | 27d | |
| Bifrost | Open source | 40 | 27d | |
| Together AI | Inference provider | 40 | 27d | |
| Helicone | Managed gateway | 43 | 27d | |
| Portkey | Managed gateway | 43 | 27d | |
| LiteLLM | Open source | 44 | 27d | |
| Orq.ai Router | Managed gateway | 45 | 27d | |
| Requesty | Managed marketplace | 48 | 27d | |
| OpenRouter | Managed marketplace | 49 | 27d | |
| Vercel AI Gateway | Managed gateway | 52 | 27d | |
| New API | Open source | 44 | 23d | |
| agentgateway | Open source | 45 | 23d | |
| MLflow AI Gateway | Open source | 45 | 23d | |
| Higress | Open source | 47 | 23d | |
| Eden AI | Managed marketplace | 54 | 23d | |
| Envoy AI Gateway | Open source | 66 | 23d | |
| Merge Gateway | Managed gateway | 130 | 23d | |
| Hugging Face Inference Providers | Managed marketplace | 88 | 22d | |
| Respan | Managed gateway | 176 | 10d | |
| AI Gateway HQ | Managed gateway | 170 | 8d | |
| Velokey | Managed marketplace | 100 | 6d |