Two numbers on every card, because they tell different stories. Web share is consumer traffic (Similarweb, September 2026). Enterprise API share is production spend (Menlo Ventures, December 2025, the latest published). ChatGPT dominates the first; Anthropic leads the second. If you only look at one, you will misread the market.
1.2B weekly users
$2–$10 / M (Sol)
Gated frontier (Daybreak)
Reasoning leader
Current lineup
GPT-6 Astra$10 / $50Sep 3. First OpenAI model self-designated “Critical” for cyber under its Preparedness Framework. Leads LiveBench Reasoning (92.65) and ARC-AGI-2 (95%). Rolled out through tiered Daybreak access. Fast mode $20 / $100.
GPT-6.1 Sol$2 / $10Sep 29 (DevDay). The balanced default. Cheapest top-tier model per task ($0.72 on the Artificial Analysis index) but coding ranks outside the top ten.
GPT-6 Luna$0.10 / $0.50High-volume tier. Cheapest tracked model per task ($0.07). Powers the new Decisions API.
GPT-6.1 Astra—Cancelled Sep 28 after internal tests showed it acting beyond scope, reaching for tools unsafely, and misreporting what it had done.
Strengths
Still the consumer default: 1.2B weekly users, 57.5% of web traffic, share rising every month since May
GPT-6 Astra leads pure reasoning benchmarks and powers the “dots” agent product
Widest price ladder in the industry: $0.10 to $50 per million output tokens inside one API
Ultrafast mode (up to 300 tok/s) and the Decisions API point at new workload types
Life-sciences specialist (gpt-rosalind-research) for approved research orgs
Weaknesses / gaps
Enterprise API share fell to 27% while Anthropic climbed to 40%
Coding lost its edge: Sol sits outside the top ten; Anthropic holds ~54% of coding spend
Safety turbulence: GPT-6.1 Astra cancelled; training of the most capable models paused pending new safeguards
UK AISI tests with safeguards off: Astra completed an unsanctioned supply-chain attack in 29.2% of runs
Frontier access now depends on tier qualification, not just payment
Strongest fits
Consumer-scale assistantsDeep reasoningHigh-volume decisioning (Luna)Marketing & content
tap to expand
950M Gemini app MAU
$2–$10 / M (intro)
Gated frontier (Fairwind)
Multimodal leader
Current lineup
Gemini 4 Argon$2 / $10 → $4 / $20 Dec 1Announced Sep 30, cyber defenders first via Fairwind (650+ partners), broader access rolling out through October. Up to 1M output tokens per response. Lowest hallucination rate measured (15%) but answers only half of questions. First on the Vals Index (finance, coding, legal, tax).
Gemini 3.8 Flash$0.75 / $3.75 → doubles Jan 1GA Sep 2. 1M input / 65K output. Introductory pricing runs through Dec 31, 2026.
Gemini Omni 1.1 Flashfrom $0.03 / secLeads Arena text-to-video (1516).
Strengths
Best multimodal stack: video generation, long-video understanding (LVBench 91.7%), image, audio
Distribution nobody can match: Search AI Mode in 19 countries, Workspace, Android, Chrome
Argon tops AutomationBench (end-to-end business processes) and the Vals Index for finance and legal work
CodeMender vulnerability-fixing agent ships alongside Argon for defenders
Enterprise share tripled since 2023 (7% to 21%)
Weaknesses / gaps
Consumer momentum stalled: share peaked at 28.6% in June and has fallen three months running
Introductory pricing cliffs: Argon doubles Dec 1, Flash doubles Jan 1. Budget accordingly.
Argon’s conservatism cuts both ways: lowest hallucination, but it declines half the time
Phased, trust-based access makes Argon hard to plan a production launch around
Many benchmark claims come from Google’s own harnesses
Strongest fits
Multimodal & videoGoogle Workspace orgsBusiness-process automationFinance & tax analysis
tap to expand
1B+ monthly visits
$0.10–$50 / M
Gated frontier (Glasswing / Mythos)
Coding & agents leader
Current lineup
Claude Opus 5.5$4 / $20Sep 22. #1 on the Artificial Analysis Intelligence Index (57.6). Leads LiveBench Agentic Coding (71.7) and WebDev Arena. 1M context, 128K output, cache reads $0.20.
Claude Sonnet 5.5$2 / $10Sep 28. #2 overall (56.0). Leads LiveBench Coding (91.4). Default model on the free tier.
Claude Haiku 5.5$0.10 / $0.50Oct 7. 75–90% cheaper than Haiku 4.5. 1M input, 128K output, adaptive thinking via an effort parameter. $0.50 / $2.50 above 100K context.
Claude Fable 5.1$10 / $50Highest factual accuracy measured on AA-Omniscience (67%). The public-access sibling of Mythos.
Claude Mythos 5.1$10 / $50Sep 1. Same specs and price as Fable 5.1 but verification-gated (e.g., Cyber Verification Program). Available on the Claude API, Bedrock, Google Cloud, Microsoft Foundry. Mythos Preview (Project Glasswing) lists at $25 / $125.
Strengths
Enterprise leader: 40% of production API spend, ~54% of the $4B coding segment
Holds the top three positions on the Artificial Analysis index (Opus 5.5, Sonnet 5.5, Fable 5.1)
Fastest-growing consumer surface: 1B+ monthly visits, +545% YoY, October tracking +17.6% over September
MCP is now an open standard under the Linux Foundation’s Agentic AI Foundation, adopted by Copilot, ChatGPT, and Gemini
Haiku 5.5 reset the price floor without giving up 1M context
Weaknesses / gaps
Consumer share still single digits (9.8%); no OS, browser, or search distribution of its own
Lived through the only government-ordered access suspension so far (June 12–30, 2026)
Top tier is slow by design: Mythos/Fable default to high effort and “slower” latency
No native video or image generation; multimodal is input-side (text + image)
Verification gating adds procurement friction for the most capable model
Strongest fits
Software engineering & agentsRegulated enterpriseLegal & researchCyber defense (verified orgs)
tap to expand
Merged into SpaceX (Feb 2)
$2 / $6 (Grok 4.7)
Real-time X data
Current lineup
Grok 4.7$2 / $6Sep 21. xAI’s current coding and knowledge model. Keeps Grok 4.6’s price and speed but emits roughly twice the output tokens, so real cost per task is higher ($2.73 on the AA index).
Grok 4.1 Fast$0.20 / $0.502M context. The volume tier.
Grok Imagine Video 1.5 / Voice 3.0—October releases. Video with synchronized sound; voice flagship Oct 14.
Strengths
SpaceX–xAI merger (finalized Feb 2, ~$1.25T combined) solves the capital problem: a profitable parent with $15–16B revenue
Only lab with native, real-time access to X’s social graph and conversation stream
Competitive mid-tier price ($2 / $6) and a 2M-context fast tier
Pushing hard on multimodal: Imagine video, Voice 3.0
Fewer refusals than peers for creative and edge-case work
Weaknesses / gaps
Consumer traffic down 37% in six months; DeepSeek now has more web share
No meaningful enterprise API footprint in published share data
Benchmark position mid-pack (AA 46.3), well behind the top three labs
Merger with a U.S. government space contractor invites regulatory scrutiny; brand volatility persists
Thin third-party ecosystem compared with OpenAI, Google, Anthropic
Strongest fits
Social & media intelligenceBrand monitoringCreative video & voice
tap to expand
The open-weight story flipped twice in six months. Meta split into open Llama and closed Muse. Chinese labs (Moonshot, DeepSeek, Alibaba, Z.AI, MiniMax, Xiaomi) now hold the top open-weight positions, with Kimi K3.1 Pro at #4 overall. And yet open weights fell to 11% of enterprise API spend as closed models pulled ahead. Capability and adoption are moving in opposite directions.
Dual-track: closed frontier + open weights
April 8: Llama 5 (open weights, up to 5M-token context, near-parity with GPT-5 on MMLU-Pro) and Muse Spark (closed, built by Meta Superintelligence Labs) shipped the same day. October 12: Muse Wonder ($4 / $20) and subscriber-only Wonder Full, with “Active Memory” and “Active Focus” for agentic tasks; Wonder Full posts the highest GDPval-AA score tracked (1857). October 15: Muse Spark Open, an 816B / 42B-active open model ($1.25 / $4.25) that runs on Apple silicon with ~1TB unified memory. Meta AI passed 1B monthly users; 2026 capex guidance is $115–135B.
Best for
Self-hosted deployments at every scale, very long documents (5M ctx), on-device and Apple-silicon inference, organizations already inside Meta’s ad and messaging ecosystem
Risks / gaps
Strategy whiplash (open, then closed, then open again) makes roadmap bets hard; Muse Wonder Full is not on the public API; Llama community license is not OSI-approved; enterprise API share remains small
tap to expand
The top open-weight model, #4 overall
Kimi K3 (July 16): 2.8T-parameter MoE, 16 of 896 experts active, 1M context, native vision, weights published 11 days after launch under an MIT-based “Kimi K3 License” (commercial agreement required above $20M revenue). Kimi K3.1 Pro (Oct 23): Artificial Analysis score 55, fourth overall and the highest of any open model; Terminal-Bench 4.0 65.7%; $2 / $10; 104B active parameters; a Highspeed variant runs 2–3x faster. Full weights due Nov 4.
Best for
Agentic coding and terminal work, self-hosted frontier-class capability, teams that want weights plus a competitive hosted API
Risks / gaps
China-origin data-handling and procurement concerns for regulated U.S. buyers; 594GB–1.4TB weights need serious infrastructure; license is not OSI open source; API pricing (K3 at $3 / $15) is not cheap
tap to expand
Frontier coding at commodity prices
V4 (April 24): Pro at 1.6T total / 49B active, Flash at 284B / 13B, 1M input and 384K output, trained on 32T+ tokens with FP4 experts. LiveCodeBench 93.5%, SWE-bench Verified 80.6%, Codeforces 3206. Pro $1.74 / $3.48, Flash $0.14 / $0.28. V4.1-Flash-1017 (Oct 17): 552B / 8B active, DeepSWE 77.0 (DeepSeek’s table puts it above Opus 5), MIT weights the same day, $0.15 / $0.60 off-peak and $0.30 / $1.20 peak, cached input effectively ~$0.017. 3.5% of consumer web traffic, now ahead of Grok.
Best for
High-volume coding agents, math and competitive programming, cost-sensitive production inference, self-hosting under a true MIT license
Risks / gaps
Geopolitical and data-residency concerns; benchmark claims are self-reported; trails on 1M-context retrieval (MRCR); peak/off-peak pricing complicates budgeting
tap to expand
Multilingual breadth and the cheapest capable API
Qwen3.8-Flash-Next is the newest downloadable Qwen; Qwen3.8-Max is a 2.4T-parameter MoE. Qwen3.7 Flash lists at $0.03 / $0.13 per million tokens, the lowest price of any widely used model. Qwen3.6 27B shipped Oct 21; Qwen Image 2.1 Turbo Oct 9. Qwen 4 was announced Sep 22 and is still in training. Apache 2.0 licensing on most open releases.
Best for
Multilingual deployments, RAG and summarization at massive scale, image generation, any workload where $0.03 input changes the business case
Risks / gaps
Qwen 4 slipped; smaller Western developer community; compliance concerns for U.S.-regulated industries; many flagship (Max) models are not openly priced
tap to expand
Europe’s sovereign frontier option
Mistral Large 4 (preview Oct 6): 1T total / 49B active, natively multimodal, 160+ languages, guardrailed Studio-only preview with weights promised by month end. Listed at $0.68 / $2.09 with 1M context. Small 4 at $0.15 / $0.60; Medium 3.5 (128B) at $1.50 / $7.50; Ministral 3 family from $0.10. The only non-U.S., non-Chinese lab shipping trillion-parameter models.
Best for
EU-regulated industries, GDPR and data-residency requirements, air-gapped government deployments, European-language NLP, edge inference (Ministral)
Risks / gaps
Large 4 is still a preview; benchmark standing behind the top U.S. and Chinese labs; smaller ecosystem; enterprise API share bundled into the “other” 12%
tap to expand
Fast iteration, 1M context, open releases
GLM-5.2 lists at $1.40 / $4.40 with 1M context; GLM-5.3 Fast shipped Oct 7 and GLM-5.3 weights are listed free. One of the four Chinese labs now shipping on a roughly monthly cadence.
Best for
Teams hedging across multiple open-weight providers, Chinese-language workloads, cost-sensitive agents
Risks / gaps
Limited Western enterprise track record; same geopolitical considerations as DeepSeek and Qwen
tap to expand
Multimodal input at $0.15 / $0.60
M3.1 Flash (API Oct 19, weights Oct 23): 428B / 23B active, 1M context, text + image + video input, always-on thinking, DeepSWE 73.7, $0.22 per task on the AA index. One caution from the same evaluation: 92.8% of its wrong answers were delivered confidently.
Best for
Cheap multimodal pipelines, video-understanding at scale, agent workloads with tight cost ceilings
Risks / gaps
Confident-wrong behavior demands guardrails and verification layers; China-origin considerations
tap to expand
Grok-class scores at a tenth of the price
MiMo-V2.6-Pro scores 46.3 on the AA index, the same as Grok 4.7, at $0.435 / $0.87 (about $0.13 per task). V2.6-Flash lists at $0.14 / $0.28. Open weights, MIT license, 1M context.
Best for
Tool-calling and coding agents at high throughput, cost-sensitive production, device-adjacent deployments
Risks / gaps
Young enterprise track record; limited ecosystem outside China
tap to expand
Fading
The cautionary tale
Perplexity’s traffic has fallen every month since March 2026 and its U.S. visitors were flat for the year before that. Citation-backed answers stopped being a differentiator once Google’s AI Mode and ChatGPT added them, and the market shifted toward agents that execute tasks. The pivot is enterprise agents, “Perplexity Computer” (long-running cloud agent), and a local “Personal Computer” device. Microsoft Copilot’s spring surge reversed by July. The lesson for buyers: a product built on someone else’s frontier model needs a moat that is not the model.
Still useful for
Fast cited research, journalists and analysts who want source links by default
Watch
Publisher and Reddit litigation; dependence on competitors’ models; whether the agent products find adoption
tap to expand
Financial services is running the table. A March 2026 PYMNTS survey of $1B+ U.S. enterprises found financial services and insurance at high adoption on 27 of 75 AI-supported tasks (media 16, healthcare 10), with 85% planning budget increases. Pick models by vertical fit, not leaderboard position.
Top use cases are revenue recognition and close (65% of firms), credit risk scoring (60%), and sales forecasting (60%). Gemini 4 Argon ranks first on the Vals Index for finance and tax work. Anthropic leads enterprise API spend and ships the lowest-latency path to verified-org models for fraud and cyber teams. GPT-6 Luna and the Decisions API fit high-volume scoring.
Factual reliability is the buying criterion. Fable 5.1 posts the highest measured accuracy on AA-Omniscience (67%); Gemini 4 Argon has the lowest hallucination rate (15%) at the cost of declining half of questions, which is often the right trade in clinical settings. OpenAI’s gpt-rosalind-research is a life-sciences specialist restricted to approved research organizations. Google’s Fairwind program prioritizes healthcare operators as critical infrastructure.
Long-context reasoning with low hallucination. Fable 5.1 for accuracy-critical review; Opus 5.5 for multi-step analysis with 1M context; Llama 5 (up to 5M tokens, self-hosted) when client data cannot leave the building. Gemini 4 Argon also ranks first on the Vals Index legal track. Watch the EU AI Act Omnibus: high-risk deadlines moved to Dec 2027 / Aug 2028, but transparency obligations still land Aug 2, 2026.
Coding is the $4B enterprise use case (55% of departmental AI spend), and Anthropic holds ~54% of it. Opus 5.5 leads agentic coding (LiveBench 71.7, WebDev Arena); Sonnet 5.5 leads raw coding (91.4). Kimi K3.1 Pro is the open-weight leader on Terminal-Bench (65.7%). DeepSeek V4.1 Flash delivers DeepSWE 77.0 at $0.30 / $1.20. Note GPT-6.1 Sol now sits outside the coding top ten.
New this edition. All three leading labs now ship their most capable models to verified defenders first: Anthropic’s Mythos via Project Glasswing and the Cyber Verification Program, Google’s Argon via Fairwind (governments, critical infrastructure, large platforms; runs without cyber guardrails for admitted teams; CWE-bench 68%, tied first), OpenAI’s Astra via Daybreak Access and Daybreak Blue. If you run a SOC, vulnerability program, or pen-test team, qualifying for these programs is now a procurement workstream.
Gemini Omni 1.1 leads text-to-video and Argon can return 1M output tokens in a single response (whole campaigns, not snippets). ChatGPT Images 2.5 holds the top two image-board positions. Meta’s Muse is the native choice inside Meta’s ad ecosystem. Grok Imagine Video 1.5 and Voice 3.0 are credible for short-form social.
Mistral Large 4 (1T params, 160+ languages, weights promised) is the only trillion-parameter model from a European lab. Llama 5 open weights for air-gapped deployment. Gateways with EU residency and on-prem options (Orq.ai, TrueFoundry, Kong) matter as much as the model. The June 2026 U.S. export-control episode is now a standing argument for sovereign capacity.
Qwen remains the breadth leader with the cheapest API ($0.03 / $0.13 for 3.7 Flash). Mistral Large 4 covers 160+ languages natively multimodal. Kimi K3 adds a 1M-context open option. For China-market deployments, note 700M+ domestic generative AI users and a separate regulatory regime.
Grok still owns real-time X data and is the only option for live social sentiment at the model layer. Gemini inherits Google’s search index and AI Mode distribution in 19 countries. Perplexity’s decline removed what used to be the third option here.
The teams building the control plane. Sonnet 5.5 and Haiku 5.5 (or GPT-6 Luna) cover the workhorse and volume tiers; what matters more is the gateway: routing, budgets, guardrails, MCP/A2A governance, and now identity. With Palo Alto Networks folding Portkey into Prisma AIRS and Kong governing agent-to-agent traffic, the gateway moved from a developer tool to a security control. See Routing & Governance.
Different roles evaluate AI through different lenses. A VP of Engineering, a CISO, a CFO, and a CMO are each right about what matters, and they are rarely looking at the same thing. Map the conversation to the buyer.
Cares about agentic reliability, cost per task, and vendor concentration. Opus 5.5 leads agentic coding; Sol is the cheapest top-tier model per task ($0.72); Kimi K3.1 Pro gives a frontier-class fallback you can self-host. Evaluates on: cost per completed task (not per token), 1M-context behavior, tool-use reliability, MCP/AGENTS.md support, and whether a second provider can take production traffic tomorrow.
The role that changed most this year. CISOs now have privileged access channels to the strongest models (Glasswing/verification, Fairwind, Daybreak) and a new control surface: the AI gateway. Three things to evaluate: (1) which verification programs your org qualifies for, (2) whether your gateway enforces identity, guardrails, DLP, and audit on every model call (Prisma AIRS/Portkey, Kong, Cloudflare), and (3) supply-chain integrity of the gateway itself after the LiteLLM PyPI compromise. The June export-control suspension belongs in your continuity plan.
New this edition. Enterprise generative AI spend hit $37B in 2025 (3.2x year over year), yet McKinsey’s 2026 survey finds only ~37% of organizations report any EBIT contribution and ~6% are high performers. The CFO lens: tier your workloads (see Cost Tiers), insist on cache-hit and batch economics, and watch introductory-pricing cliffs (Gemini Argon doubles Dec 1, Flash doubles Jan 1). Evaluates on: cost per task, contract escalators, and measurable P&L attribution.
Content at scale across modalities. Gemini Omni for video, Argon for long-form output (1M tokens per response), ChatGPT Images 2.5 for stills, Muse inside Meta’s ad stack. Evaluates on: creative quality, multimodal breadth, brand-safety guardrails, and integration with the existing martech stack.
Weights or it didn’t happen. Kimi K3 (2.8T MoE), DeepSeek V4 Pro (MIT, FP4-trained), Llama 5 (5M ctx), Muse Spark Open (runs on Apple silicon), Qwen 3.8, GLM-5.3, MiniMax M3.1, Mistral Large 4 (weights promised). Evaluates on: license terms (several are MIT-based with revenue clauses, not OSI), fine-tuning flexibility, memory footprint (Kimi K3 is 594GB at MXFP4), and benchmark reproducibility.
Day-to-day coding and the agent toolchain. Sonnet 5.5 leads LiveBench Coding and is the free-tier default; DeepSeek V4.1 Flash is the cheap coding agent; IDE-native models like Cursor Composer 2.5 ($0.50 / $2.50) are now real options. Standards consolidated: MCP, AGENTS.md, and goose live under the Agentic AI Foundation. Evaluates on: code quality, IDE integration, agentic tool-use, and standards support.
Accuracy and auditability. Fable 5.1 for factual precision, Opus 5.5 for multi-step reasoning, Argon when a non-answer beats a wrong answer. Regulatory calendar: EU AI Act GPAI obligations in force since Aug 2025; transparency obligations Aug 2, 2026; high-risk Annex III moved to Dec 2, 2027 and Annex I to Aug 2, 2028 under the Digital Omnibus (provisional agreement May 6, 2026). Evaluates on: hallucination rate, citation behavior, data residency, and audit trail.
Synthesis with sources and judgment. Opus 5.5 and Sol for board-level synthesis across long documents; Fable 5.1 when the facts have to be right. Two things to keep in perspective: “AGI era” language in launch materials is positioning, not measurement; and consumer share and enterprise share are telling opposite stories, so ask which one a vendor is quoting.
How token pricing works: models charge per million tokens (roughly 750K words), quoted as input / output. Output costs more because it takes more compute. The cost driver is rarely the answer; it is the context you send with every request. Three levers now matter as much as list price: cache reads (90–95% off repeated input across every major lab), batch APIs (50% off async work), and introductory pricing cliffs (several flagship prices double on a fixed date).
Mythos Preview: $25 / $125 (Project Glasswing) · GPT-6 Astra: $10 / $50, fast mode $20 / $100 (Daybreak tiers) · Fable 5.1 / Mythos 5.1: $10 / $50 (Mythos verification-gated)
Access depends on qualification, not just budget. Use for cyber defense, the hardest reasoning, and accuracy-critical review. Opus 5.5 costs $5.98 per task on the AA index; these cost more and run slower by design.
Claude Opus 5.5: $4 / $20 (#1 overall) · Meta Muse Wonder: $4 / $20 · Gemini 4 Argon: $4 / $20 from Dec 1 (currently $2 / $10)
The default for agentic work that has to be right the first time. Note Opus 5.5 is 20% cheaper than Opus 4.6 was in April while adding 1M context.
Claude Sonnet 5.5: $2 / $10 · GPT-6.1 Sol: $2 / $10 · Gemini 4 Argon (intro): $2 / $10 · Kimi K3.1 Pro: $2 / $10 · Grok 4.7: $2 / $6 · Kimi K3: $3 / $15
Where 80% of production should live. Four different labs now converge on $2 / $10, which makes multi-provider routing nearly free to set up. Watch output volume: Grok 4.7 emits ~2x the tokens of its predecessor, so its real cost per task ($2.73) exceeds Sol’s ($0.72).
Gemini 3.8 Flash: $0.75 / $3.75 (doubles Jan 1, 2027) · Mistral Large 4: $0.68 / $2.09 · MiMo-V2.6-Pro: $0.435 / $0.87 · DeepSeek V4.1 Flash: $0.30 / $1.20 peak, $0.15 / $0.60 off-peak · MiniMax M3.1 Flash: $0.15 / $0.60 · Mistral Small 4: $0.15 / $0.60
Chatbots, classification, extraction, bulk summarization. Most of this tier is open-weight, so the hosted price is a ceiling, not a floor.
Claude Haiku 5.5: $0.10 / $0.50 (1M ctx; $0.50 / $2.50 above 100K) · GPT-6 Luna: $0.10 / $0.50 ($0.07 per task, cheapest tracked) · DeepSeek V4 Flash: $0.14 / $0.28 · MiMo-V2.6-Flash: $0.14 / $0.28 · Qwen3.7 Flash: $0.03 / $0.13
This tier did not exist in April. The budget floor fell roughly 10x in six months (Haiku 4.5 was $1 / $5). Routing, triage, decision calls, and anything you run millions of times a day belongs here.
Kimi K3 (2.8T, 594GB at MXFP4) · DeepSeek V4 Pro / V4.1 Flash (MIT) · Llama 5 (community license, 5M ctx) · Muse Spark Open (816B / 42B, Apple silicon ~1TB) · Qwen 3.8, GLM-5.3, MiniMax M3.1 (open) · Mistral Large 4 (weights promised)
Frontier-class capability is now downloadable, but the infrastructure bar rose with it: trillion-parameter MoEs need multi-node serving. Read the license: several “open” models carry revenue thresholds or attribution clauses and are not OSI-approved.
⚡
The 70 / 20 / 10 rule, re-tiered
Route ~70% of traffic to the near-free tier, ~20% to a workhorse, ~10% to a flagship, and escalate only on failure. In April the budget tier was $1 / $5; today it is $0.10 / $0.50, so the same blend costs a fraction of what it did. Stack cache reads (up to 95% off input) and batch (50% off) on top. Budget realistically at ~1.5–2x modeled token cost for growth and experimentation, and put every introductory price on a calendar.
Intro-pricing cliffs on the calendar: Gemini 4 Argon $2 / $10 → $4 / $20 on Dec 1, 2026. Gemini 3.8 Flash $0.75 / $3.75 → $1.50 / $7.50 on Jan 1, 2027. DeepSeek charges 2x at peak hours. Kimi K3.1 Pro full weights arrive Nov 4, which may move its hosted price.
The AI gateway stopped being a developer convenience and became a security control. It is the one place every model call passes through, which makes it the natural enforcement point for identity, budgets, guardrails, DLP, and audit. Two acquisitions in 2026 made that official.
Disclosure: I work at Palo Alto Networks, which acquired Portkey in 2026. This guide is my own independent analysis, built only from public sources, and is not reviewed or endorsed by my employer. Vendors here are listed on public merits; weigh them accordingly.
Routing
Classify by task, send to the cheapest tier that passes, escalate on failure; :floor / :nitro style variants; peak/off-peak awareness
Resilience
Provider failover, retries, and now a plan for a model vanishing by government order (18 days in June)
Cost control
Virtual keys, per-team budgets, semantic cache, showback / chargeback, intro-price tracking
Security & governance
Agent identity, RBAC, PII / DLP, input and output guardrails, MCP and A2A traffic policy, full audit trail
What changed
The gateway becomes part of a security platform
Portkey (founded 2023, ~$18M raised, 3,000+ models through one API, the deepest built-in guardrail catalog, MIT-licensed core, ~13K GitHub stars) is being integrated into Prisma AIRS as its in-path enforcement point: routing agent requests, applying policy, authenticating agents, with the agent registry tied to identity capabilities from the CyberArk acquisition ($25B, closed February 2026). A Gateway 2.0 pre-release merges the enterprise gateway into the open-source build. Analyst read: AI governance is shifting from a software tool category to a core security platform requirement, and the market is consolidating around a few control planes.
tap to expand
Your gateway holds your provider keys
LiteLLM (~53.8K stars, 100+ providers) survived a PyPI supply-chain compromise in March. The lesson is structural: a gateway sits in the path of every model call and holds every provider credential, so it needs the same supply-chain discipline as any identity system: pinned versions, hash verification, signed images, and an owner. Helicone’s move into maintenance mode under Mintlify is a reminder that single-purpose observability gateways are getting absorbed.
tap to expand
The field
Fastest open gateway
~8.7K stars, 23+ providers through one OpenAI-compatible API. Claims 11µs added latency at 5,000 RPS; Maxim’s own benchmark shows P99 of 1.68s vs 90.72s for LiteLLM at 500 RPS. Virtual keys with hierarchical budgets in the OSS build, fallbacks, semantic caching, MCP gateway, Prometheus and OpenTelemetry. Enterprise pricing is custom.
tap to expand
Broadest provider coverage
Chat, embeddings, images, audio, batches, rerank, MCP and A2A. Virtual keys, per-team spend tracking, RPM/TPM limits, admin UI, guardrails, Langfuse and OpenTelemetry logging, Terraform modules, cosign-signed Docker images (post-incident). 8ms P95 at 1K RPS by its own benchmarks.
tap to expand
Agent-to-agent governance
Extends Kong’s API gateway to LLM, MCP, and agent-to-agent traffic. PII sanitization, semantic prompt guards, access control, user/model/time-bound token quotas with showback and chargeback, A2A telemetry. Gartner API-management leader for the seventh year. Best fit for shops already standardized on Kong.
tap to expand
Free DLP on every plan
One-line setup on Cloudflare’s edge. Retries, model fallback, free DLP scanning, Llama Guard guardrails (billed as Workers AI), Unified Billing with no inference markup. Core features (analytics, caching, rate limiting) free on all plans. Cloudflare also now ships its own decision models (Clef, Apache 2.0).
tap to expand
Air-gapped and on-prem
1,600+ models, under 3ms internal overhead (vendor claim), VPC / on-prem / air-gapped / multi-cloud, latency- and geo-aware routing, self-hosted model serving (vLLM, SGLang, KServe, Triton), SOC 2, HIPAA, GDPR, SSO, RBAC, audit logs. Free for the first 30K requests per month. 4.8 on Gartner Peer Insights (210 ratings).
tap to expand
Fastest path to multi-model
Vercel AI Gateway: callable from any infrastructure, covers text through video and realtime, budgets per team/project/key, no token markup including BYOK. OpenRouter: hosted OpenAI-compatible access to every major lab, :nitro and :floor routing variants, 5.5% fee on credits, provider prices passed through. Orq.ai Router: 400+ models, routing tied to evals and budgets, SOC 2, GDPR, EU data residency.
tap to expand
⚖
The control plane is the product
Models are converging on price ($2 / $10 at four labs) and on context (1M everywhere). What is not converging is governance: who can call which model, with what data, at what cost, under whose identity, with what audit trail. The gateway is where those answers get enforced, which is why security platforms are buying them and why Gartner, Kong, and the Linux Foundation are all building for agent-to-agent traffic. If you have more than one model in production and no gateway, that is the gap.