What each case runs. Low — the smallest models available (Ministral 3B 70%, Ministral 8B 20%, Llama 4 Scout 10%), memory held on self-managed storage, 5% of payload sampled to logging; approx. 10% accuracy, 3.7s latency. Base — a middle mix (Ministral 8B 50%, Llama 4 Scout 25%, Qwen3 Next 80B 15%, imported SLM 10%) on the managed memory service, 5% sampled; approx. 22% accuracy, 5.0s. High — GPT-5.6 Luna at $0.20/$1.20 per million tokens, asked once, full payload capture; 80% accuracy, 3.2s. Max — the same model asked three times per task with a fourth pass reading all three candidates and picking the winner, which multiplies tokens by 3.35 and raises accuracy by roughly 8 points to 88%, at 6.5s because the three samples run in parallel and only the verifier is serial. High and Max therefore differ in method rather than in model, and the same is true of Akka: 80.2% asking once, 88% asking three times. OpenAI cut Luna 80% from $1.00/$6.00 on 30 July 2026, AWS passed it through on Bedrock the same day and Microsoft matched it at parity on Azure from 1 August. Google Vertex carries no OpenAI proprietary models, so both tiers there run its strongest open weight instead, reaching 73% asked once and 81% asked three times. OpenAI published no SWE-bench Verified score for the GPT-5.6 family; published figures are SWE-bench Pro 62.7%, Terminal-Bench 2.1 84.7% and a Coding Agent Index of 75. On SWE-bench Pro, Luna leads DeepSeek V4 Pro at 55.4%, and V4 Pro scores 80.6% on Verified, so the 80% single-pass placement is reasoned rather than measured. Luna runs on AWS-managed infrastructure with in-region processing and the model provider has no access to prompts or completions, so its data-residency posture matches an open-weight model on the same platform; what a proprietary model cannot do is carry a trained adapter, run on capacity you reserve, or get cheaper as your own traffic accumulates.
Akka — MiniMax M2.5 and Qwen3 Coder Next with customer-trained LoRA adapters, served in a single pass on reserved capacity; 80.2% accuracy, 4.5s. Best-of-three sampling is not applied, so the Akka figure is the published score of the base model rather than a modelled uplift.
Model inference. Input tokens × input rate plus output tokens × output rate, summed across each mix. Rates per million tokens, Amazon Bedrock public list, on-demand, US East: Ministral 3B $0.10/$0.10; Ministral 8B $0.15/$0.15; Llama 4 Scout $0.20/$1.00; Qwen3 Next 80B $0.15/$1.20; Mistral Large 3 $0.50/$1.50; Qwen3 VL 235B $0.53/$2.66; DeepSeek V3.2 $0.62/$1.85; MiniMax M2.5 $0.30/$1.20, corroborated by one published rate card, with a second showing $0.36/$1.44, exactly 20% higher, which is consistent with a gateway uplift rather than a differing Bedrock rate. The Amazon Bedrock model card for MiniMax M2.5 records a launch date of 12 February 2026, a 196K context window and an 8K maximum output, and refers pricing to the Bedrock pricing page. Best-of-three sampling applies a 3.35× multiplier: each attempt re-reads the full context, tripling both input and output, plus one verifier pass reading the context and all three candidates. Customer-trained models on Bedrock bill through Custom Model Import at $0.05718 per Custom Model Unit-minute, divided by 120,000 tokens per minute at an assumed 2,000 tokens per second per unit, giving $0.4765 per million. Amazon Bedrock documentation states that a model unit specifies the input and output tokens it can process in one minute, and directs customers to their AWS account manager for the figure rather than publishing it. The 2,000 tokens per second assumption is therefore ours, and it decides whether an imported-model strategy has a cost basis at all.
GCP basis. Vertex AI does not sell open-weight models by the token. Google models bill per token, third-party models follow vendor pricing, and open models bill per endpoint instance hour plus a per-model container fee. The inference rates here are therefore derived rather than quoted. An a4-highgpu-8g endpoint carrying eight NVIDIA B200 lists at $16.11 per GPU-hour, and is used in preference to a3-highgpu-8g at $88.49 an hour for eight H100, because three times the serving throughput for 1.46 times the rate is 51% cheaper per token and is the fairer basis for Google. Dividing that rate by the serving throughput modelled elsewhere on this page, and again by 40% endpoint utilisation since a rented endpoint bills whether or not it is serving, yields approximately $0.09/$0.62 per million tokens for a small model, $0.25/$2.39 for a mid-size model and $0.37/$4.14 for a large one. The per-model container fee is charged on top, is published per Marketplace listing rather than centrally, and is excluded here, so the Google figures understate the real cost. Reserved capacity inside a Google VPC uses the published three-year committed-use discount of 6% on a3, the only committed rate of the three hyperscalers that is published rather than assumed; a4 is excluded from that discount because Google routes A4 through its AI Hypercomputer reservation system rather than the committed-use machinery a3 uses. The Max case runs the strongest open model in Model Garden, reaching approximately 81%.
Agent runtime and memory differ by provider. These two lines are derived per hyperscaler because the billing models diverge far more than the rates. AWS bills AgentCore Runtime at $0.0895 per vCPU-hour and $0.00945 per GB-hour, and AgentCore Memory at $0.25 per 1,000 events, $0.75 per 1,000 records per month against the accumulated balance, and $0.50 per 1,000 retrievals; the monthly charge against a growing balance is why the AWS memory line compounds. Azure makes no charge to create or run an agent in Foundry Agent Service, but that service is a control plane rather than a runtime: Microsoft bills model tokens, search services, storage and compute invoked by the agent. Agent code and tools are modelled on Azure Container Apps at $0.000024 per vCPU-second and $0.000003 per GiB-second, which is $0.0864 per vCPU-hour and $0.0108 per GiB-hour, plus $0.40 per million requests. That makes Azure compute dearer than AgentCore Runtime, not free. Memory is modelled at $670,000 a year covering Azure Cosmos DB for conversation history, Azure Storage for files, and Azure AI Search for vector retrieval at S3 pricing of $981 per search unit per month, which bills for every hour a unit exists regardless of query volume. That remains an infrastructure comparison rather than a capability one: AgentCore Memory performs extraction and consolidation that a customer assembling Cosmos DB and AI Search would have to build. Google bills Vertex AI Agent Engine at $0.0864 per vCPU-hour and $0.0090 per GB-hour, and sessions and Memory Bank events at $0.25 per 1,000 with no monthly charge against stored records, so its memory line does not compound. Observability, storage, support and professional services carry the AWS figures on every provider: Google Cloud Logging bills $0.50 per GiB with thirty days of storage included against CloudWatch Logs at $0.50 per GB plus $0.03 per GB-month, and Google Premium Support bills the greater of $15,000 a month or 10% to $150,000, 7% to $500,000, 5% to $1,000,000 and 3% above, identical to AWS Enterprise Support. Akka charges no fee per agent execution on any provider.
Azure basis. Azure AI Foundry serverless rates per million tokens: Ministral 3B $0.04/$0.04, Phi-4 $0.125/$0.50, Llama 4 Scout $0.20/$0.78, Mistral Medium 2505 $0.40/$2.00, Mistral Large 3 $0.50/$1.50, DeepSeek V3.2 $0.58/$1.68. Azure prices at or below Bedrock on every model both platforms carry. The mixes differ from the AWS ones because Azure does not serve Ministral 8B, Qwen3 Next 80B, Qwen3 Coder Next, Qwen3 VL 235B or MiniMax M2.5. Phi-4 and Mistral Medium substitute for the Qwen and Ministral 8B positions. The Max case runs DeepSeek V3.2, the strongest model with a published Azure rate, so its ceiling is approximately 81% rather than the 88% MiniMax M2.5 reaches on Bedrock. Llama 4 Maverick is excluded: the published rate of $1.41 input against $0.35 output appears transposed. Durable memory, agent hosting, observability, storage, support and professional services carry the AWS figures, on the basis that those services price comparably across hyperscalers; only inference and reserved GPU are re-derived. Reserved capacity inside an Azure VPC is modelled at $59.00 an hour for an ND96isr H100 v5, a 40% reduction on the $98.32 on-demand rate, which is an assumption rather than a quoted figure.
Durable memory. Bedrock AgentCore. Short-term events $0.25 per 1,000; long-term storage billed monthly against the accumulated balance at $0.75 per 1,000 per month, or $0.25 per 1,000 per month self-managed; retrieval $0.50 per 1,000. Volumetrics assume 50,000 tokens per session, with 8 short-term events, 3 long-term records and 2 retrievals per session. The billed balance is the mean of a linear ramp to full year-one accrual; because stored records accumulate, year two bills approximately 2.5× year one at flat token volume. Agent hosting. AgentCore Runtime, $0.0895 per vCPU-hour billed on active execution and $0.00945 per GB-hour billed across the full session wall clock including time blocked awaiting a model response; session profiles 1 vCPU / 2 GB / 120s, 2 / 4 / 300s and 3 / 6 / 400s. Observability. Payload estimated at 4 bytes per token; CloudWatch Logs Standard ingestion $0.50 per GB, storage $0.03 per GB-month, custom metrics $0.30 per metric per month; routing payload to S3 Standard at $0.023 per GB-month approximately halves the same volume. CloudWatch does operate volume tiers of $0.50 per GB to 10 TB a month, then $0.25, $0.10 and $0.05, but those tiers apply only to AWS vended logs such as VPC Flow Logs and Route 53 query logs. Agent request and response payload reaches CloudWatch through PutLogEvents and the CloudWatch agent, which are excluded from tiering and bill flat at $0.50 per GB. Tiering is also assessed per account and cannot be aggregated across a multi-account structure. The Infrequent Access log class at $0.25 per GB has supported full Logs Insights queries since early 2026 and halves ingestion where it is acceptable.
Enterprise Support Plan. Billed as the greater of $15,000 per month or a tiered percentage of monthly consumption: 10% to $150,000, 7% to $500,000, 5% to $1,000,000, 3% thereafter; a Technical Account Manager is included at this tier at no separate charge. Professional services. Blended at $350 per hour; Low approx. 500 hours, Base two engineers for six months (1,920 hours), High and Max three to four engineers for twelve months. Availability. The AWS figure of 99.9% is the monthly uptime percentage committed in the Amazon Bedrock Service Level Agreement. The Akka figure of 99.9999% is the availability commitment of the Akka Agentic AI Platform.
Reserved capacity. Sized as continuous demand × 2.8 for peak, plus one spare node, and never below the two nodes required to hold MiniMax M2.5 and Qwen3 Coder Next resident with high availability across two availability zones. Two fleets are sized against the same work and the cheaper is taken: p5.48xlarge carrying eight NVIDIA H100 at 80GB, and p6-b200.48xlarge carrying eight NVIDIA B200 at 180GB and modelled at three times the H100 serving throughput. Below roughly 20 trillion tokens a year the H100 fleet wins, because node count is set by peak headroom rather than by throughput and the cheaper node carries the idle hours. Above that the B200 throughput advantage overtakes its higher hourly rate. Per-GPU serving throughput modelled at 25,000 tokens per second prefill and 3,000 tokens per second decode; prefill is compute-bound and decode is bound by memory bandwidth. The 2.8 peak-to-average factor is the measured figure for coding workloads across 44 million requests over nine days in the published Microsoft Azure LLM inference traces; conversational workloads in the same traces measure 1.7. Supporting compute and storage scale with deployed agents and end-user traffic at 12.5% of reserved capacity, subject to a $50,000 annual minimum. Adapter training executes against idle reserved capacity and carries no incremental charge; a trained workload consumes approximately 40% of prior input tokens and 60% of prior output tokens on a model requiring half the compute per token, and adapters are served as LoRA weights on a shared base model, adding under 1 GB of GPU memory each.
Accuracy. SWE-bench Verified, a public benchmark scoring resolution of real software defects, weighted by each mix. Published results: Llama 4 Scout 47.3%; DeepSeek V3.2 73.1%; Qwen3 Coder Next 74.2%; MiniMax M2.5 80.2%. Estimated where no result is published: Ministral 3B 5%; Ministral 8B and customer-trained models 8%; Qwen3 Next 80B 35%; Mistral Large 3 40%; Qwen3 VL 235B 55%. The 8% estimate uses Qwen3-8B, measured at 8.0%, as a size-class proxy; neither Ministral variant publishes a SWE-bench result. Best-of-three sampling with a verifier rerank is modelled as adding approximately 8 percentage points, anchored on a published result in which a 77.2% model reached 90.0% within an engineered agent scaffold. That uplift is applied only to the Max case, which pays for three samples and a verifier pass. It is not applied to Akka. Adapters trained on repeating work raise reliability on those tasks, and that gain is not claimed here because SWE-bench Verified measures general software engineering rather than the narrow tasks an adapter serves. Latency. Blended time to first token plus 500 output tokens at blended output speed, from Artificial Analysis: Ministral 3B 188.4 tokens/second at 0.64s; Ministral 8B 113.6 at 0.73s; Llama 4 Scout 119.9 at 0.83s; Qwen3 Next 80B 187.5 at 2.20s; Mistral Large 3 40.3 at 1.37s; DeepSeek V3.2 108.8 on Bedrock at 0.92s.
Three-year projection. Throughput per GPU-hour is projected to rise 1.6× in year two and 2.6× cumulatively in year three on unchanged hardware, from speculative decoding (1.30×), four-bit quantisation on tolerant traffic (1.25×), newer open-weight models activating fewer parameters (1.40×), inference engine improvements (1.15×) and additional trained adapters (1.15×), which compound to 3.00×; the published projection takes approximately half. Independently observed open-weight inference cost decline runs 30% to 50% annually.
General. All AWS amounts are public list price at on-demand rates, US East, as of August 2026. No Enterprise Discount Program, Private Pricing Agreement, Reserved Instance, Savings Plan, Capacity Block or other committed-use discount is applied; rates already held with AWS reduce these amounts. Amounts exclude customer personnel, third-party software and non-AWS infrastructure. This estimate is a planning aid. It is not a quotation and it commits neither Akka nor Amazon Web Services. Sources: Amazon Bedrock, Bedrock AgentCore, AWS Support and Amazon CloudWatch pricing pages; Amazon EC2 instance pricing; the Amazon Bedrock Service Level Agreement; Artificial Analysis; the Microsoft Azure LLM inference traces; Epoch AI inference price trends.
Scope. One use case carries the entire volume entered above. Inference and specialist training are both directed at that use case. Amounts are annual.
Base model. Each API model is matched to a base model. Where the API model is open weight, the base model is that same model, self-hosted. Where it is proprietary, the base model is the nearest open weight by capability class: Claude Fable 5 and Claude Cowork take Kimi K3 at 2.8 trillion total and 104 billion active parameters, the largest open-weight model; Claude Opus 5 and GPT-5.6 Sol take DeepSeek V4-Pro at 1.6 trillion total and 49 billion active; Claude Sonnet 5 takes Kimi K2.6 at 1 trillion total and 32 billion active; GPT-5.6 Terra takes GLM-5.2 at 744 billion total and 40 billion active; GPT-5.6 Luna takes MiniMax M2.5; Claude Haiku 4.5 takes Qwen3 Coder Next.
Parameter counts. Throughput scales on active parameters rather than total, since a mixture-of-experts model activates a subset of its weights per token. Qwen3 Coder Next activates three billion parameters. Kimi K3, DeepSeek V4-Pro, Kimi K2.6 and GLM-5.2 carry the active counts stated above. Ministral 3B, Ministral 8B, Phi-4 and Mistral Large 3 are dense, so active and total are the same.
Accuracy. A trained specialist model reaches approximately 90% on the task it is trained for. That figure applies where the use case is narrow. Breadth of the use case, rather than its subject, determines whether it is reached.
Training. With training off, the base model is self-hosted and served as published, and the comparison is the cost of running that model against the cost of calling an API. With training on, a specialist model is trained on the use case: it consumes 40% of the prior input tokens and produces 60% of the prior output tokens, on a model requiring half the compute per token. Prompt caching and speculative decoding apply in both cases. Off-peak batching applies in neither.
Token split by use case. Writing software 90% input, answering customers 80%, reading documents 95%, search and summary 92%, tidying data 85%. The API rate applied is the blend of that provider's input and output rates at this split, not its headline rate.
Fleet sizing. Continuous demand multiplied by 2.8 for peak, plus one spare node, and never below two nodes. The 2.8 peak-to-average factor is the measured figure for coding workloads across 44 million requests in the published Microsoft Azure LLM inference traces. Hyperscalers bill whole nodes of eight GPUs; neoclouds bill per GPU. The compute is the one selected in the provider list and the capacity picker above.
Akka fee. The Akka figure is the fee, which carries margin. Margin is 50% at or below five trillion tokens a year and declines on a logarithmic curve to 18.8% at one thousand trillion. Compute and storage follow a declining schedule per five trillion tokens: $100,000 for the first block, decaying toward a $15,000 floor, subject to a $50,000 annual minimum. Network and licence costs add $12,500 a year, plus $35,000 a year for private connectivity where Akka runs outside the customer's own cloud.
Per-token rates. Anthropic, OpenAI and DeepSeek first-party rates verified 16 August 2026. AWS Bedrock rates verified 12 August 2026 from the published pricing page. Azure AI Foundry rates verified 12 August 2026 from an aggregator; Azure renders its pricing pages client-side. DeepSeek rates are cache-miss input; peak and off-peak pricing was announced for 16:00 UTC on 16 August 2026. Anthropic and OpenAI models are proprietary: weights and adapters are not distributed, so a fine-tuned workload runs only on open-weight models.
Serving throughput. 25,000 tokens per second prefill and 3,000 tokens per second decode per H100 at a 12B dense equivalent, both scaling as 1/N with active parameter count, derived from floating-point throughput and memory bandwidth. Those figures are first-principles rather than benchmarked, and mixed sequence lengths are expected to reach only half to two-thirds of them, so a factor of 1.75 is applied to the Akka side throughout this tab.
Prompt caching. The input side of every per-token rate is charged at 12.07% of list. That figure is measured across 969 Claude Code transcripts and 50,279 API calls: 98.2% of context tokens were served from cache, 1.8% written to it, and a negligible share billed fresh, with cache reads billing at a tenth of the input rate and writes at 1.25 times it. Output is charged in full. Batch processing at 50% and committed-spend discounts, reported to average 28% on negotiated enterprise agreements, are available on several of these providers and are not applied. Per-outcome products bill a completed unit of work, so their price does not move with how cacheable a workload is and no cache discount is applied to them.
Per-outcome rates. Some products price a completed unit of work rather than a token. Each is converted to a per-million-token equivalent by dividing its rate by the tokens that unit consumes. Microsoft Copilot Studio bills 2 Copilot Credits a generative answer at $0.01 a credit. Salesforce Agentforce bills 20 Flex Credits an action at $500 per 100,000 credits, which is $0.10 an action; its alternative conversation model is $2.00. HubSpot Breeze bills $0.50 a resolved conversation, Aissist $0.60 a resolution, Gorgias $0.90, Intercom Fin $0.99 and Zendesk AI Agents $1.50. Claude Cowork bills $13.57 a task, reported as Claude 5 Fable High in the Creative Strategies Diligence Stack Agent Bench for financial analysis and research. Where a vendor publishes a range the lower bound is used. A resolution or conversation is converted at 50,000 tokens, an action or generative answer at 5,000, and a Cowork task at 985,000 — the median session measured across 969 Claude Code transcripts and 50,279 API calls, at 981,000 context tokens and 4,000 output tokens, running 349 input tokens to every output token with 98.2% of context served from cache. Sierra, Decagon and Ada advertise per-outcome pricing without publishing a rate. ServiceNow Now Assist bills per assist, 25 for a small generative task and 150 for a large agentic action, without a published price per assist. Workday Illuminate includes credits in the subscription without a published per-agent rate.
General. All amounts are in US dollars at list price. No enterprise discount, private pricing agreement, reserved instance, savings plan, capacity block or other committed-use discount is applied beyond those stated in the reserved capacity rates. Amounts exclude customer personnel, third-party software and infrastructure outside the fleet described. Sources: Amazon Bedrock and Amazon EC2 pricing; Azure AI Foundry and Azure Virtual Machines pricing; Google Cloud Vertex AI and Compute Engine pricing; Anthropic, OpenAI and DeepSeek platform pricing; Creative Strategies Diligence Stack Agent Bench; the Microsoft Azure LLM inference traces; Artificial Analysis.
