# AI Training Costs 2026: GPT-5.6, Claude Fable 5, Gemini > Training a frontier model costs $1B+. Breakdown of compute, energy, data, and R&D costs across the three leading labs, and what they mean for API prices. Source: https://capitalandcompute.net/blog/what-it-costs-to-train-ai-models-2026/ Published: 2026-07-12 Updated: 2026-07-31 The three most capable AI models in July 2026 cost somewhere between $2 billion and $4 billion to develop, combined. That is more than the GDP of a dozen small countries. And the number that matters most is not the total. It is how that money was spent, and what it implies about the prices you pay per token. $1.5B-$3B GPT-5 total program cost compute, R&D, data, energy $500M-$1B Claude Fable 5 estimated cost Anthropic does not disclose; extrapolated from cluster size $500M-$1B+ Gemini 3.5 estimated cost Google's largest training run to date 65-75% Compute share of total cost GPU rental dominates every lab's budget These figures come with a warning label: every lab keeps its actual training cost confidential. The numbers in this post are estimates synthesized from Epoch AI compute projections, SemiAnalysis hardware analysis, public cluster-size disclosures, energy-pricing models, and disclosed inference-cost data. They are directional, not audited. But they are the best picture we have of where a billion dollars goes when you train a frontier model. The short version * Training a flagship model costs $500M to $3B, with compute (GPU-hours) consuming 65-75% of the total. The rest is R&D labor, data acquisition, energy, and failed experiments. * Inference already costs more than training over the model lifecycle. Annual inference spend for a deployed frontier model exceeds its one-time training cost within 6-18 months of release. * The cost gap between labs is driven by architecture choices (MoE vs. dense), hardware stack (NVIDIA vs. TPU), and training efficiency. DeepSeek proved a capable model can cost under $1M with severe optimization. * API pricing does not track training cost linearly. Fable 5 charges $50/Mtok output, but its training cost is probably 3-5x lower than GPT-5.6\. The price reflects scarcity and positioning, not cost recovery. * Training cost growth is accelerating: GPT-3 cost $4.6M (2020), GPT-4 \~$78M (2023), GPT-5 $1.5B+ (2025). At this trajectory, a single training run could exceed $10B by 2029. ## The headline numbers Every frontier model is a capital project on the scale of a skyscraper or a semiconductor fab. The table below lines up the three leading labs on a consistent set of metrics, drawing on the best available public estimates. | Metric | GPT-5 (base) | GPT-5.6 (incremental) | Claude Fable 5 | Gemini 3.5 | | ----------------------- | --------------- | --------------------- | --------------- | ---------------- | | Architecture | \~3-5T MoE | \~3-5T MoE | Likely 1-2T MoE | Likely >5T MoE | | Training tokens | \~20-30T | \~5-10T (additional) | \~10-20T | \~15-25T | | Total FLOPs | \~5e25 - 1e26 | \~1e25 - 3e25 | \~1e25 - 5e25 | \~3e25 - 8e25 | | GPU cluster size | \~50K H100/B200 | \~50K H100/B200 | \~20-40K GPU | \~30-50K TPU v5p | | GPU-hours | \~100M-200M | \~30M-50M | \~40M-80M | \~60M-120M | | Peak cluster power | \~75-120 MW | \~75-120 MW | \~30-60 MW | \~50-80 MW | | Est. total cost (range) | $1.5B-$3B | $300M-$500M | $500M-$1B | $500M-$1B+ | Estimated training infrastructure for the three leading frontier models, mid-2026\. Figures are synthesized from multiple sources (see Sources section). Lab-disclosed numbers are marked; everything else is inferred from cluster size, training duration, and hardware cost. A few things jump out. GPT-5 cost more than the other two combined, partly because it was first to the new scale tier and partly because OpenAI disclosed the broadest set of figures (the others keep their cards closer). The incremental GPT-5.6 training run, which produced the Sol/Terra/Luna tier splits, cost a fraction of the base model because it started from the same weights and added RL at reduced compute. Claude Fable 5 and Gemini 3.5 are wider ranges because neither Anthropic nor Google has published training details. The lower bounds assume efficient training with high Model FLOPs Utilization (MFU) and competitive hardware pricing. The upper bounds assume a less efficient training process and retail cloud rates. __Training cost range by model__ | Item | High estimate | Low estimate | | -------------- | ------------- | ------------ | | GPT-5 (base) | $3.0B | $1.5B | | GPT-5.6 (RL) | $0.5B | $0.3B | | Claude Fable 5 | $1.0B | $0.5B | | Gemini 3.5 | $1.2B | $0.5B | Training cost range for each model. The bar spans low to high estimate; the dot marks the midpoint. GPT-5 dominates because it established the current scale tier. ## Breaking down the cost stack A billion-dollar training run is not a single line item. It is five distinct cost categories with very different economics. | Cost category | Share of total | Typical cost (for \~$2B run) | What drives it | | ----------------------------- | -------------- | ---------------------------- | ----------------------------------------------------------- | | GPU compute | 65-75% | $800M-$1.3B | Hardware rental, cluster interconnects, idle time | | R&D and engineering | 15-20% | $250M-$400M | Research scientist salaries, architecture design, alignment | | Data acquisition and curation | 5-10% | $80M-$150M | Licensing, synthetic data generation, filtering, dedup | | Energy | 2-6% | $40M-$80M | Electricity at $0.04-0.08/kWh, PUE 1.2-1.4 | | Infrastructure overhead | 3-5% | $50M-$100M | Networking, storage, cooling, facility colo | Estimated cost breakdown for a \~$2B frontier training run. Compute (GPU-hours) is the dominant line item at 65-75%. Failed experiments add 20-30% on top of the final run. How this was made These percentages are derived from Epoch AI’s cost-modeling framework and cross-checked against public statements from OpenAI (Altman’s “most expensive model ever built” comments), SemiAnalysis data-center cost models, and GPUnex’s provider-price survey. The energy figure assumes a 100-day training run at 80MW average draw with PUE 1.3 and US industrial power rates. Labs with access to cheaper power (Google’s renewable PPAs) or better cooling will land at the low end. ### GPU compute: the 800-pound item The single most important number in AI economics is the cost of one GPU-hour. At market rates in mid-2026: * **H100 SXM** on a specialized cloud: $2.00-$3.00/hr * **H100 SXM** on a hyperscaler: $3.50-$4.50/hr * **B200** (Blackwell): $4.00-$6.00/hr * **TPU v5p** (estimated internal cost): $1.50-$2.50/hr Fifty thousand GPUs running for 100 days at 80% utilization generates roughly 96 million GPU-hours. At a blended $4/hr (mix of H100 and B200 on hyperscaler rates), that is $384M. Add cluster interconnects (InfiniBand or NVLink, typically 10-15% of compute cost) and the total lands at $440M-$500M for the single successful training run. That is before the failed runs. Frontier training is not a single shot. Labs report that 2-3 failed or partially successful training runs precede the final one, each burning a fraction of the final run’s compute. With two failed runs at 30% of final each, the total compute cost climbs to \~$700M. For a comparison of what those GPU-hours cost at different tiers of hardware, see our [self-hosted LLM cost-per-token analysis](/blog/self-hosted-llm-cost-per-token/). ### Energy: small share, strategic leverage Energy accounts for only 2-6% of total training cost, but it is the hardest line item to compress. An 80MW cluster draws 76,800 MWh over 100 days. At $0.055/kWh (typical US industrial rate with PUE 1.3), that is $4.2M per month, or roughly $14M for the full run. Double it for the failed runs: $28M. The strategic leverage comes from location. A lab that builds its data center in a region with $0.025/kWh power (Quebec, Scandinavia, parts of the US Pacific Northwest) cuts energy cost by more than half. Google, which runs the world’s largest corporate renewable energy procurement, likely pays materially less per MWh than OpenAI or Anthropic for equivalent clusters. ### Failed experiments: the hidden multiplier This is the cost that almost nobody talks about. The public “training cost” of GPT-5 ($1.5B-$3B) includes far more than the single successful training run. Epoch AI estimates that frontier labs spend 40-60% of their total compute budget on experiments that do not produce a shippable model architecture, data mixes, learning rate schedules, and scaling-law validation runs that inform the final design but consume GPU-hours with no revenue attached. When you read that GPT-5 cost $2B, roughly $800M of that was for runs that produced no model. That is not waste. It is the R&D cost of discovering the training recipe. But it means the effective cost of the shipped model is a fraction of the total program cost. ## Why costs vary so much between labs The four models in our comparison span a roughly 5x cost range despite targeting similar capability levels. The spread comes down to four variables. ### Architecture efficiency The single biggest cost lever is Model FLOPs Utilization (MFU), the fraction of theoretical peak compute that the training software actually achieves. A lab running at 55% MFU gets through training in half the wall-clock time of one running at 30% MFU, cutting GPU-hours proportionally. OpenAI and Google both invest heavily in custom training infrastructure (OpenAI’s supercomputing partnership with Microsoft, Google’s TPU-native stack). Anthropic, which relies primarily on third-party cloud GPU capacity, may face lower MFU due to less optimized interconnects and scheduler overhead. ### Hardware stack NVIDIA GPUs (H100, B200) dominate the training market, but they carry a hefty margin. Google’s TPUs, while not available for rent externally, cost Google less per FLOP because the hardware is internal and the software stack is purpose-built. This gives Google a structural cost advantage of roughly 30-40% on compute for equivalent training runs. __Relative compute cost per unit of training: Google vs. Anthropic__ | Metric | Google (TPU advantage) | Anthropic (cloud GPU) | | ----------------------- | ---------------------- | --------------------- | | GPU/TPU cost per FLOP | 100% | 100% | | Interconnect cost | 70% | 130% | | Energy cost per FLOP | 60% | 110% | | Software MFU | 120% | 80% | | Est. total compute cost | 70% | 120% | Estimated hardware cost per effective FLOP for each lab's training stack. Google's TPU vertical integration gives it a structural cost advantage on raw compute. Anthropic, as a cloud GPU renter, pays the most per FLOP despite likely running smaller clusters. ### Location and energy costs Google’s data center portfolio includes facilities in Finland, Belgium, Taiwan, and the US Pacific Northwest, all regions with below-average industrial electricity prices. OpenAI’s primary training infrastructure (leased Microsoft Azure clusters) is distributed across Virginia, Iowa, and Arizona, where power costs 40-80% higher than Google’s lowest-cost sites. Over a 100-day training run, this difference alone could add $10M-$20M to OpenAI’s energy bill. ### The DeepSeek counterexample DeepSeek V3, trained for a reported $5.6M in compute (later refined to \~$6M including data and engineering), and DeepSeek R1 at \~$294K, proved that frontier-competitive models can be trained on budgets three orders of magnitude below the US labs. The caveats are significant: DeepSeek used export-restricted H800 GPUs at below-market rates, optimized aggressively for a single known architecture, and likely does not include the full R&D cost of the earlier experiments that enabled the efficient final run. Even so, the DeepSeek numbers show that training cost is not a fixed function of capability. It is a function of how much you are willing to spend to get there faster. ## What training cost means for API pricing This is the question that matters to anyone using these models. If GPT-5 cost $2B to train, why did GPT-5.6 Luna cost the same per token as Gemini 3.1 Pro, which cost a quarter as much to train, and why could OpenAI then cut Luna by 80 percent overnight on July 30, 2026? Because API pricing has almost nothing to do with training cost recovery. Insight API prices are set by what the market will bear and by competitive positioning, not by amortized training cost. A model that cost $2B to train can charge the same per-token rate as a model that cost $500M, because pricing is a function of perceived value and willingness to pay, not cost-plus margin. The real cost structure for inference is far simpler. Once a model is trained, serving it costs: * **GPU compute for inference:** $0.50-$5.00 per million output tokens (depending on model size, quantization, batching, and hardware) * **Energy for inference:** 5-10% of compute cost * **Amortized training cost over total inference tokens:** negligible for any model that serves more than a few billion tokens At GPT-5.6 scale, OpenAI serves trillions of tokens per month across its API. Even a $2B training cost amortized over two years of inference at, say, 10 trillion tokens per month adds roughly $0.008 per million tokens, a rounding error next to the $30/Mtok output price. __Training amortization vs. inference cost vs. API price ($/Mtok)__ | Item | High | Low | | ------------------------------------ | ------ | ------ | | GPT-5.6 training (annualized 2yr) | $0.010 | $0.005 | | GPT-5.6 inference compute (per Mtok) | $5 | $1 | | GPT-5.6 API price (Sol output) | $30 | $30 | | Fable 5 inference compute (per Mtok) | $8 | $2 | | Fable 5 API price (output) | $50 | $50 | Training cost, annualized, is a tiny fraction of the per-token price for any model deployed at scale. Inference cost dominates. API prices reflect inference cost plus margin, not training cost recovery. The small bar at the top is the training amortization. The middle bar is the inference cost. The large bar is the API price. The gap between inference cost and API price is margin for the provider, not training cost recovery. Caution A note on the numbers: inference cost varies enormously with batching strategy, quantization, and hardware. The ranges above assume FP8 inference on H100-class hardware with moderate dynamic batching. A provider using INT4 quantization on purpose-built inference chips (Groq LPU, Google TPU) or serving at extremely high batch sizes could cut inference cost by 3-5x, widening the margin further. The same cost dynamics apply beyond training: the token price on your API dashboard is often the smallest line item in your actual bill. See our analysis of the [hidden cost of AI-generated code](/blog/hidden-cost-ai-generated-code/) for a breakdown of where the real money goes. ## The trajectory that cannot continue Training costs have been growing at roughly 2.4x per year since 2016, a rate that Epoch AI tracks and projects forward. If that trend holds: * **GPT-3** (2020): $4.6M * **GPT-4** (2023): \~$78M * **GPT-5** (2025): $1.5B-$3B * **2027-2028**: $5B-$10B per training run * **2029-2030**: $10B-$25B per training run These numbers quickly become unsustainable. No single product generates enough revenue to amortize a $10B training cost in a reasonable timeframe. The number of organizations that can afford frontier training shrinks from dozens (today) to single digits (2028) to perhaps two (2030). Something has to give. The most likely escape valves: 1. **Algorithmic efficiency improvements** that decouple capability from compute cost (DeepSeek’s 100x efficiency ratio is a preview) 2. **Hardware specialization** that continues to drive down FLOP cost (NVIDIA’s annual architecture cadence, inference-only chips) 3. **Model distillation at scale**, where the expensive training cost is incurred once and the capability is distributed through cheap student models 4. **Consolidation**, where only two labs (likely OpenAI and Google) can afford frontier training, and everyone else fine-tunes or distills Each of these is already happening. The question is whether their combined effect can offset the 2.4x/year cost growth. | Model | Year | Est. training cost | Notes | | -------------- | ---- | ------------------ | --------------------------------------------------------- | | GPT-3 | 2020 | \~$4.6M | OpenAI disclosed; \~1,000 V100 GPUs, 34 days | | PaLM | 2022 | \~$12M | Google; 6,144 TPU v4, 50 days | | GPT-4 | 2023 | \~$78M | Compute only; \~10K A100, \~100 days | | Gemini Ultra | 2024 | \~$191M | Compute only; 16K TPU v5, \~130 days | | Llama 3.1 405B | 2024 | \~$60M | Meta disclosed; 16K H100, \~54 days | | DeepSeek V3 | 2025 | \~$5.6M | Compute only, aggressive optimization; 2K H800, \~55 days | | GPT-5 | 2025 | $1.5B-$3B | Total program, including R&D and failed runs | | Claude Fable 5 | 2026 | $500M-$1B | Estimated from disclosed cluster capacity | Estimated training cost per flagship model, 2020-2026, with trend projection to 2029\. Points are logged on the y-axis; the slope represents \~2.4x annual growth. The DeepSeek V3 point shows the efficiency frontier. The table makes visible a tension that defines the industry. The cost of frontier training has climbed from single-digit millions to billions in six years. But at the same time, the cost of a capable model for any specific task has collapsed. DeepSeek V3 delivered a model competitive with GPT-4 at 1/14 the compute cost, three years later. Inference prices have fallen 1,000x since 2023. The effective cost of AI capability is dropping for everyone except the handful of labs racing at the very frontier. For builders, operators, and end users, the trend is benign. Prices will continue to fall. Capability will continue to rise. The $2B training run is someone else’s problem. ## Frequently asked questions ## Frequently asked questions Why do training cost estimates vary so widely? Labs do not disclose their actual training costs. Every public estimate is inferred from cluster size, training duration, hardware rental rates, and disclosed FLOP counts. Different analysts make different assumptions about MFU, GPU pricing (retail vs. negotiated bulk rate), and whether to include failed runs and R&D labor. The range from $1.5B to $3B for GPT-5 reflects this uncertainty. Does the training cost affect what I pay for API access? Almost never. API prices are set by inference cost, competitive positioning, and willingness to pay. Training cost is amortized over trillions of tokens and contributes less than $0.01/Mtok for any model deployed at scale. If your API bill is high, the cause is the volume of tokens you use, not recouping the training cost. How did DeepSeek train a competitive model for $5.6M? DeepSeek V3 used three main efficiency levers: a Mixture-of-Experts architecture with sparse activation, aggressive FP8 training that cut memory and compute in half, and a training recipe optimized through extensive earlier experiments (which themselves cost more than the final run). It also used export-restricted H800 GPUs at below-market rates. The $5.6M figure covers only the final training run, not the full R&D program. Will training costs keep growing at 2.4x per year? Not indefinitely. The trend is already bending as the industry shifts from scaling parameters to scaling inference-time compute (the o1/o3 reasoning paradigm). The next frontier may be RL-based post-training, which shifts the cost from pre-training FLOPs to inference-time thinking tokens. This could cap pre-training cost growth at current levels while capability continues to improve. Who can afford to train frontier models in 2026? Realistically, OpenAI, Google, Anthropic, and Meta. Microsoft has the capital but has chosen to partner with OpenAI rather than train its own frontier model. xAI, DeepSeek, and Alibaba (Qwen) can train at the tier below frontier. By 2028, the list may narrow to two. Does training a model on TPUs cost more or less than GPUs? For the lab that owns the TPUs, it costs significantly less per FLOP because there is no hardware margin baked in. Google likely pays 30-50% less per FLOP for TPU training than OpenAI pays for GPU training on Azure. However, the comparison only works if the software stack is mature enough to achieve high MFU on TPUs, which is a nontrivial engineering investment. ## Sources * Epoch AI (2025). _Notes on GPT-5 Training Compute_. Epoch AI Substack. * Epoch AI (2024). _Trends in Training Compute_. * SemiAnalysis (2025). _The Economics of AI Training Clusters_. SemiAnalysis. * GPUnex Research Team (2026). _How Much Does It Cost to Train an AI Model in 2026?_ GPUnex Blog. * BtMData (2026). _GPT-5: Training Infrastructure Breakdown_. * Anthropic (2026). _Introducing Claude Opus 4.6 / Fable 5_. Anthropic News. * Google DeepMind (2024). _Gemini: A Family of Highly Capable Multimodal Models_. arXiv. * Stanford HAI (2025). _AI Index Report 2025_. * About Chromebooks (2026). _Machine Learning Model Training Cost Statistics 2026_. * Trixly AI Solutions (2025). _GPT-5: The Most Expensive AI Model Ever Built_. * GridStack Hub (2026). _AI Infrastructure Costs Explained: GPU, Network, Storage, Energy (2026)_.