Skip to content

Model Landscape — April 2026

The four axes that define every model: open vs closed, frontier vs mid-tier vs small, native multimodal vs text-first, reasoning vs instant. The 2026 plot twist: the gap between open and closed has nearly closed. DeepSeek V4, GLM-5.1, Qwen 3.5 are within a few points of GPT-5.4 / Claude Opus 4.7 on most benchmarks at 1/20 to 1/50 the price.

Rapid Recall

Closed frontier: GPT-5.4 (unified routing, OSWorld 75%, \(2.50/\)15), Claude Opus 4.7 (coding, Constitutional AI), Gemini 3.1 Pro (GPQA 94.3%, cheapest frontier at \(2/\)12), Grok 4 (HLE 50.7%). Open-weight frontier: DeepSeek V4 (1T MoE, \(0.28/\)0.50, ~50× cheaper), Llama 4 Scout (10M context), Qwen 3.5, Mistral, GLM-5.1, MiniMax M2.7, Kimi K2.5. India: Sarvam-M outperforms DeepSeek R1 on some benchmarks at 1/6 size; Saaras V3 ASR beats Gemini 3 Pro and GPT-4o on IndicVoices. 2026 reality: routing 70% Sonnet + 20% Haiku + 10% Opus replaces "Opus for everything" at 60%+ cost reduction. API prices dropped ~80% from 2025 to 2026.

§1 The four axes

Four axes define every model.

  • Open vs closed (can you download weights and self-host?).
  • Frontier vs mid-tier vs small (capability tier).
  • Native multimodal vs text-first (audio/video/image as first-class inputs?).
  • Reasoning vs instant (explicit chain-of-thought with test-time compute vs fast single-pass).

The "gap between open and closed has nearly closed" in 2026. Six months ago, top closed models held a commanding lead. As of April 2026, open-weight models from Chinese labs (DeepSeek V4, GLM-5.1, MiniMax M2.7, Kimi K2.5) and Meta (Llama 4) are within a few points on most benchmarks at 1/20 to 1/50 the price.

§2 Closed / proprietary frontier

2.1 OpenAI — GPT-5.4 (released March 2026)

  • Replaces GPT-5.3 Codex as flagship.
  • Key feature: unified routing architecture — the model itself decides fast-reply vs deep-reasoning based on query complexity.
  • 1M token context in Codex mode, native computer use, 47% tool-call token reduction.
  • Pricing: $2.50 / $15 per M tokens (standard), $30 / $180 (Pro with max reasoning).
  • Best at: OSWorld computer use (75%, above human expert), terminal/DevOps, structured reasoning.
  • GPT-4o is still around as the cheaper multimodal workhorse but is legacy now. When people say "GPT-4o" in 2026 they usually mean "the realtime voice/audio model."

2.2 Anthropic — Claude Opus 4.7 / Sonnet 4.6 / Haiku 4.5

  • Claude Opus 4.7 (most recent): top of Anthropic's line, highest reasoning depth and code quality.
  • Sonnet 4.6: ~98% of Opus quality at a fraction of the cost — the workhorse most production apps use.
  • Haiku 4.5: cheap/fast tier for classification, simple extraction.
  • Pricing (Opus 4.6 as reference): $5 / $25 per M tokens, 1M token context.
  • Best at: coding (powers Cursor, Windsurf, Claude Code), nuanced writing with natural prose, long-context coherence, complex instruction following.
  • Distinctive: trained with Constitutional AI (see Constitutional AI and RLAIF) rather than pure RLHF.

2.3 Google DeepMind — Gemini 3.1 Pro (released Feb 19, 2026)

  • Leads 12 of 18 tracked benchmarks.
  • Pricing: $2 / $12 per M tokens — cheapest frontier model.
  • Best at: scientific reasoning (GPQA Diamond 94.3%), abstract reasoning (ARC-AGI-2 77.1%, double predecessor), native multimodal across text/image/audio/video, 1M context.
  • Gemini 2.5 Deep Think is the reasoning-specific variant.
  • Gemini 3 Flash is the cheap/fast tier.

2.4 xAI — Grok 4 / Grok 4.1 Fast

  • Leads HLE (Humanity's Last Exam) at 50.7% — the hardest reasoning benchmark.
  • Native integration with X/Twitter real-time data.
  • Less censored on sensitive topics (deliberate positioning).
  • Pricing: $2 / $15 per M tokens.

§3 Open-weight frontier (the 2026 story)

3.1 DeepSeek V4 / DeepSeek-R1

  • V4 released March 2026, trained on Huawei Ascend chips (no NVIDIA at all).
  • 1T parameters MoE, $0.28 / $0.50 per M tokens — ~50× cheaper than GPT-5.4.
  • R1 is their reasoning model — uses pure RL with GRPO (see GRPO), no SFT cold start in R1-Zero variant.
  • OpenAI publicly accused DeepSeek in Feb 2026 of distilling from GPT outputs (see Distillation).
  • Known limits: censorship on politically sensitive topics (Chinese regulation), sometimes slow due to long CoT.

3.2 Meta — Llama 4 (Scout / Maverick / Behemoth)

  • Scout: 10M token context window (largest in industry), 17B active params MoE.
  • Maverick: higher capability tier.
  • Behemoth: reportedly in training, not released as of April 2026.
  • Fully open-weight with commercial license, strong for self-hosting.
  • Real-world coding performance has been disputed — benchmarks vs actual developer use show a gap.

3.3 Alibaba — Qwen 3.5

  • Multilingual strength (Chinese, Arabic, Hindi, Spanish).
  • Efficient MoE (397B total, 17B active).
  • Sonnet-class quality runnable on 2× RTX 4090.
  • Ultra-cheap API ($0.11 / 1M input).
  • Distilled variants (0.8B, 2B, 4B) dominate the small-model leaderboard for latency / cost.

3.4 Mistral, Microsoft Phi-4, others

  • Mistral (French lab): Mistral Small 3, Magistral, Codestral. European alternative, GDPR-friendly.
  • Phi-4 (Microsoft): ~14B params competitive with 70B on reasoning, trained on synthetic textbook-quality data. Great for edge.
  • Other Chinese frontier open-weight worth knowing: GLM-5 / GLM-5.1 (Z.ai, 94% of Claude Opus 4.6 coding), MiniMax M2.5 / M2.7 (80.2% SWE-bench), Kimi K2 / K2.5 Thinking (Moonshot), Baichuan, Yi, Hunyuan.

§4 India-specific: Sarvam AI

The one set of models you should be able to speak about fluently as a Hyderabad-based AI PM. Sarvam AI (Bengaluru, backed by IndiaAI Mission).

  • Sarvam-M (multilingual reasoning LLM): 30B and 105B variants, launched Feb 18, 2026. The 105B model outperforms DeepSeek R1 (600B) on several benchmarks at 1/6 the size. 22 scheduled Indian languages plus English, 128K context, efficient MoE (128 experts, top-6 / top-8 routing).
  • Saaras V3 (ASR): outperforms Gemini 3 Pro and GPT-4o on IndicVoices and Svarah benchmarks for Indian languages and Indian-accented English. Streaming real-time, trained on 1M+ hours of multilingual audio, optimized for code-mixing and noisy telephony.
  • Bulbul (TTS for Indian languages).
  • Saarika (speech-to-text, transcription-focused).
  • Mayura (text translation).
  • Sarvam Vision (OCR + document understanding, 93.28% on OmniDoc).
  • Sarvam Audio: speech-to-command directly (skips the ASR → LLM → TTS cascade).
  • Edge models running on feature phones (Nokia/HMD partnership), Bosch cars, Sarvam Kaze smart glasses.

Why this matters for India: 1.45B people, most prefer voice in regional languages. Sarvam has a structural moat — global models cannot match it on Indian languages because training data, speaker diversity, and code-mixing are deeply local.

§5 The 2026 decision tree

What is your task?
├── Need data sovereignty (HIPAA, GDPR, India PDP)?
│   → Self-host Llama 4 / Qwen 3.5 / DeepSeek V4
│   → Or Sarvam-M for Indian language apps
├── Need maximum reasoning?
│   → GPT-5.4 Pro, Claude Opus 4.7 with thinking,
│     Gemini 2.5 Deep Think, or DeepSeek-R1 if budget-constrained
├── Need cheap high-volume?
│   → Gemini 3.1 Pro ($2/$12), Haiku 4.5,
│     or open-weight on vLLM (DeepSeek V4, Qwen 3.5)
├── Need real-time voice?
│   → gpt-realtime (OpenAI), Gemini Live,
│     Sarvam Saaras V3 for Indian languages
├── Need long context (1M+ tokens)?
│   → Gemini 3.1 Pro, Claude Opus 4.7, Llama 4 Scout (10M)
├── Need coding agent?
│   → Claude Opus 4.7 (or Sonnet 4.6 for cost),
│     GPT-5.4, GLM-5.1 if open-weight required
└── Need small/edge?
    → Phi-4, Qwen 3.5 4B, Gemma 3n, or Sarvam edge models

§6 The 2026 pricing collapse

One stat to anchor every cost conversation: API prices dropped ~80% from 2025 to 2026. What cost $500/month in mid-2025 runs $50-100 today.

  • Input tokens: $0.11 to $30 per M (50× range).
  • Output tokens: 3-8× input pricing (median 4×).
  • Cached input: 50-90% discount if you hit the cache.

This is why routing systems now pay back in weeks. A workload that uses Opus for everything at \(25/M output can route 70% to Sonnet (\)15/M) and 20% to Haiku ($4/M), keeping 10% on Opus, and drop the bill by 60%+ with negligible quality loss. See Caching and routing.

§7 April 2026 frontier pricing (per M tokens)

Model Input Output Context Notes
GPT-5.4 $2.50 $15 272K / 1M Pro variant \(30/\)180
GPT-5.4 Pro $30 $180 1M Max reasoning
Claude Opus 4.7 $15 $75 1M Top Anthropic
Claude Sonnet 4.6 $3 $15 1M 98% Opus quality
Claude Haiku 4.5 $0.25 $1.25 200K Cheap tier
Gemini 3.1 Pro $2 $12 1M Cheapest frontier
Grok 4 $2 $15 256K HLE leader
DeepSeek V4 $0.28 $0.50 128K 50× cheaper
Qwen 3.5 Plus $0.11 $0.33 128K Cheapest competent
Llama 4 Maverick (self-host) GPU cost GPU cost 1M Open-weight
Sarvam-M 105B Indian rates — 128K Indian-language specialist

§8 The "what should I use" quick picks

  • Default chat app → Claude Sonnet 4.6 (quality/cost sweet spot).
  • Code generation → Claude Opus 4.7, Sonnet 4.6 for cost, or GLM-5.1 open-weight.
  • Graduate-level reasoning → Gemini 3.1 Pro or GPT-5.4 Pro.
  • Math olympiad → DeepSeek-R1 or Gemini 2.5 Deep Think.
  • Real-time voice (English) → OpenAI gpt-realtime.
  • Real-time voice (Indian languages) → Sarvam cascade (Saaras + Sarvam-M + Bulbul).
  • Cheap classification/extraction → Haiku 4.5, Gemini Flash, or Qwen 3.5 small.
  • Long-document analysis → Gemini 3.1 Pro or Claude Opus 4.7 (1M context).
  • Agent with tool use → GPT-5.4 (best OSWorld) or Opus 4.7.
  • Self-host for data sovereignty → Qwen 3.5 or Llama 4 on vLLM.

§9 Post-training recipe stack (who uses what)

Lab Recipe Signature technique
OpenAI (GPT-5.x) SFT + RLHF (PPO) + RLAIF + RLVR for reasoning Unified router, CoT compression
Anthropic (Claude) SFT + Constitutional AI + RLAIF Self-critique, constitution
Google (Gemini) SFT + RLHF + RLAIF + multimodal RLVR Deep Think mode
DeepSeek SFT + GRPO + RLVR Pure-RL reasoning (R1-Zero), rule-based rewards
Meta (Llama 4) SFT + DPO + RLAIF Open-weight release
Alibaba (Qwen 3.5) SFT + DPO + GRPO MoE efficiency
Sarvam SFT on Indian data + DPO + domain RLHF Multilingual from-scratch pretraining

Interview Questions

Q1: Walk me through the 2026 model landscape and how you would choose one for a new product.

Four axes: open vs closed, capability tier, multimodal vs text, reasoning vs instant. For a new product in 2026, I would build a routing architecture rather than pick one model. Cheap queries (intent classification, simple extraction) go to Haiku 4.5 or Gemini Flash at under $1/M tokens. Medium queries go to Sonnet 4.6 or Gemini 3.1 Pro. Complex reasoning escalates to Opus 4.7, GPT-5.4 Pro, or DeepSeek-R1 for cost-sensitive cases. If data sovereignty matters, self-host Llama 4 or Qwen 3.5 via vLLM. For Indian-language voice workloads, Sarvam's stack is non-negotiable. The decision is 50% about cost-per-task, 30% about latency budget, 20% about data/compliance.

Q2: What does the open/closed gap "nearly closing" actually mean operationally?

Six months ago, top closed models (GPT-5.2, Gemini 3 Pro, Claude Opus 4.5) held a commanding lead. As of April 2026, open-weight models from Chinese labs (DeepSeek V4, GLM-5.1, MiniMax M2.7, Kimi K2.5) and Meta (Llama 4) are within a few points on most benchmarks at 1/20 to 1/50 the price. Operationally: for data-sovereignty workloads or cost-sensitive verticals, self-hosting an open-weight model on vLLM is now a real production option, not just a cost-savings experiment. The closed-model premium is justified only for the very hardest reasoning and tool-use tasks.

Q3: Why is routing the winning architecture and not a single best model?

Cost. Output token pricing spans 50× across providers (Qwen 3.5 at $0.33 to GPT-5.4 Pro at $180). Most user queries are simple and can be answered by Haiku or Flash; a small fraction need Opus or GPT-5.4 Pro. Routing captures that distribution: send simple queries to cheap models, escalate when complexity demands. Reported savings 50-80% vs "Opus for everything." Plus, routing systems are easier to evolve as new models drop into the catalog.

Q4: What is Sarvam's structural advantage and why can global frontier models not catch up?

Three things. Training data: Sarvam has access to deeply local Indian-language data (22 scheduled languages, code-mixing, telephony noise, regional accents) that is not in the web crawls global labs use. Speaker diversity: Saaras V3's training set has 1M+ hours of Indian-accent audio. Code-mixing: "Haan ji, mera account number 12345 hai, I want to check balance" is normal Indian speech; global models trained mostly on monolingual English do not handle this naturally. Result: Sarvam-M outperforms DeepSeek R1 (~6× bigger) on Indian benchmarks; Saaras V3 beats Gemini 3 Pro and GPT-4o on IndicVoices.

Q5: How would you reduce a company's AI API costs by 50% in 2026?

Two levers in order. First, model routing: send simple queries to Haiku 4.5 or Gemini Flash, escalate to Sonnet 4.6 or GPT-5, reserve Opus and GPT-5.4 Pro for genuinely hard cases. Typically captures 50 to 70% savings on its own. Second, semantic caching: compute embeddings of incoming queries, match against a vector cache, serve cached responses when similarity exceeds threshold. Customer-support workloads often hit 40 to 70% cache rates. Together, 70 to 80% cost reduction is realistic. Tertiary optimizations: prompt caching at the provider level (50 to 90% off on cached prefixes), moving high-volume workloads to self-hosted vLLM with quantized open-weight models. See Caching and routing.