Amazon Bedrock vs Fireworks AI Inference

Amazon Bedrock 98

Model reading. IAM with SigV4, roles and STS session credentials, policies down to a model or inference-profile ARN, short-term API keys that expire within 12 hours, long-term keys with expiry and reset, and condition keys that can deny bearer tokens altogether (30). The security page says customer data isn't shared with model providers or used to improve the base models (20). The abuse detection page says Bedrock stores no inputs or outputs by default, and a per-Region none mode blocks models that require retention, but Claude Fable 5 and 5.1 keep all traffic up to 30 days and GPT-6 and GPT-5.x keep flagged traffic up to 30 days (13 of 15, a judgement call; Anthropic's own API got 12 for the same Fable exception without an account-level block). CloudTrail logs API calls, and model invocation logging can send prompts and responses to S3 or CloudWatch Logs (15). Bedrock is in ISO, SOC and CSA STAR Level 2 scope, HIPAA eligible and FedRAMP High in GovCloud (US-West) per its security page, with a vulnerability disclosure programme on HackerOne, and the aws.amazon.com security.txt, expired on 6 October, was renewed with Expires 2027-10-16 when read on 6 November (20). Up from 96 on 6 October for that reason.

Fireworks AI Inference 77

Read by the HTTP API checklist, as for cerebras-inference and together-inference. Service accounts with API keys issued on their behalf, roles assigned at creation and changeable (30). An inference-user role can call models but not create or modify resources (20). No inference prompt-injection mitigation established (0). Enterprise audit logs with a REST API, and cost exports per API key (15). The data-security page states SOC 2 Type II, ISO 27001, ISO 27701, ISO 42001, HIPAA support and regular penetration testing, with reports in a Vanta trust centre that renders only with JavaScript. fireworks.ai/.well-known/security.txt answered 404 on 1 November, as on 19 October, and no disclosure policy or bug bounty was found. That matches people-data-labs (12) and launchdarkly (12), which state SOC 2 Type II and ISO 27001 with no security.txt, disclosure policy or bounty, so the 20 the 19 October check gave is corrected (12). The terms (section 3.6) rule out training on customer content and inference has zero retention by default, recorded for reviewers and not scored on this checklist. 30+20+0+15+12 = 77.

Amazon Bedrock 81

Model reading. The last model or API change is 30 October 2026 (OpenAI model API support and cross-Region inference in GovCloud), after structured outputs for Claude on both endpoints on 18 October (30). Models launched since 7 September 2026 get a 6-month or 45-day Legacy period, but the legacy page, which promises at least 6 months in Legacy, lists Claude 3 Haiku, 3 Sonnet, 3.5 Sonnet, 3.5 Sonnet v2 and 3.7 Sonnet as Legacy from 10 or 13 October with end of life on 13 November in several Regions, about 34 days (6 of 12, down from 10). Six models reached end of life in the 90 days to 6 November on five dates, Command R and R+ on 19 August, Claude 3 Haiku on 10 September, Nova Sonic on 14 September, Nova Canvas and Nova Reel on 30 September and Claude Sonnet 4 on 14 October (3 of 8, down from 4). Dated document history several times a month, What's New and re:Post, with direct support on a paid plan (12 of 15). SDK issue replies not sampled (5 of 10). Current official SDKs, boto3 1.43.125 on 5 November and @aws-sdk/client-bedrock-runtime 3.1163.0 (15). Supported runtimes stated, Python 3.10 or later and Node 20 or later (10). Down from 86 on 6 October.

Fireworks AI Inference 65

Read by the HTTP API checklist, as for cerebras-inference and together-inference. Rune 26B-A4B joined serverless on 27 October, five days before this check (30). Dated changelog entries several times a month, among them 27, 21, 19 and 18 October (20). Support replies not tested (0). Official Python SDK current, v1.2.30 tagged on 21 October after eleven tags from 1 October, no official TypeScript SDK (15). Dependency health not audited (0). The serverless overview promises 'at least 2 weeks advance notice' before a model is removed, and the 27 August removals were posted the day they took effect (see the deduction). 30+20+0+15+0 = 65.

Amazon Bedrock 85

Closed service under the AWS Customer Agreement and Service Terms, with the SDKs and Smithy models under Apache-2.0 (18, as for S3). The data retention, abuse detection and security pages agree that Bedrock stores nothing by default, name the models that retain traffic and for how long, and say retained data isn't shared with providers. The retention page's default mode says AWS 'may retain the data for safety and abuse-prevention purposes' without a period, and a legacy provider_data_share mode is kept though sharing isn't supported (25 of 30). A written lifecycle policy with Legacy and EOL dates per model, but the legacy table's 34-day Legacy periods for five Claude models in some Regions contradict its own 6-month floor (15 of 20, down from 18). AWS's sub-processor list, per the 4 October S3 check, and per-Region, geographic and global inference options, with global routing able to process data in any commercial Region (18 of 20).

Fireworks AI Inference 69

Proprietary service, Apache-2.0 Python SDK. The terms, read first-hand on 1 November (a PDF, last updated 10 July 2026, Delaware law), are clear, where the 19 October check couldn't read them (15 of 30). The data-handling page and section 3.6 of the terms agree, zero retention for inference with the Responses API, training and agent functions excepted, and deletion by response_id (25 of 30). Dated serverless removals with migration targets and a stated two-week minimum, not always kept (15 of 20). Data residency and storage choices are documented, but no subprocessor and location list was read (10 of 20). 15+25+15+10 = 65.

Amazon Bedrock 92

Model reading of the checklist (tool use, structured output, caching, context, batch, SDKs, errors). Converse takes tools with toolChoice auto, any or a named tool and strict: true, but forced choice depends on the model behind it (17 of 20). Structured outputs on supported models only (13 of 15). Explicit and implicit prompt caching, cache reads at $0.20 per million on Opus 5.5, 0.05x input, with 5-minute and 1-hour TTLs (15). 1M-token context on Opus 5.5, Sonnet 5.5 and Nova 2 Lite, a judgement call since Bedrock has no default model (15). Batch at half price on supported models, but the pricing page shows none for Opus 5.5 or Sonnet 5.5, while Flex is half price on Nova (7 of 10). AWS SDKs in every major language plus token generators for Python, JavaScript and Java (10). Typed errors with recovery advice per code (15).

Fireworks AI Inference 80

Read by the HTTP API checklist, as for cerebras-inference and together-inference. Compact model API with bounded output, max_tokens and streaming (25). Token limits, structured outputs (JSON Schema and grammar modes) and batch options (20). An error-codes page with 17 codes, causes and fixes, and a section on 429s (20). Retry guidance without a general deduplication guarantee (5 of 20). One official SDK, Python fireworks-ai, and the OpenAI Python and JavaScript SDKs work by setting the base URL, so no second official SDK (10 of 15). Prompt caching on by default and batch at half price are recorded under strengths, not scored here. 25+20+20+5+10 = 80.

Amazon Bedrock 80

AWS Health Dashboard with per-service, per-Region feeds and a public history file (20). The history lists Amazon Bedrock as disrupted in ME-CENTRAL-1 (UAE) and ME-SOUTH-1 (Bahrain) since March 2026, where AWS said on 15 September 2026 (still inside the 90-day window on 6 November) that it can't restore resources held only there. No other Bedrock event appears in the history in the 90 days to 6 November; an EU-SOUTH-2 packet-loss event on 4 October named Bedrock AgentCore, not Bedrock. This counts as one major event (10). StatusGator's third-party mirror shows 'Increased Error Rates' warnings on 24, 26 and 27 August and 10 September that we couldn't find in AWS's own history, so we didn't score them. Per-model token quotas with default values in the AWS General Reference, such as 30,000,000 tokens a minute for Opus 5.5 global inference (15). The error guide gives exponential backoff with jitter for 429, 503 and 529 and says to honour Retry-After, and the SDKs retry ModelNotReadyException up to five times (15). SLA of 99.9 per cent a month per Region for the model APIs, last updated 4 October 2023 (10). Converse and the current Claude, GPT-6 and Nova 2 Lite models are GA (10).

Fireworks AI Inference 70

Public status page at status.fireworks.ai with a component per serverless model (20). The history page, read on 1 November, lists 236 incidents from 3 August to 31 October, nearly all short degradations of a single serverless model. Two hit many serverless models at once. On 11 August a cloud provider outage was 'affecting many of our serverless and dedicated deployments', with five serverless components (GPT OSS 120B, MiniMax M3 and the three Kimi K3 endpoints) in partial outage from 17:15 to 20:55 UTC and again from 01:36 to 01:49 UTC on 12 August, resolved 13 August. On 14 October a cloud provider outage set 18 serverless components to degraded performance, GPT OSS 120B, Kimi K3 and GLM 5.3 among them, from 23:06 UTC, with requests to them that 'may fail', recovery 'across the board' at 00:59 UTC on 15 October and the incident resolved at 01:53. Each is an hour or more of errors across most of the serverless catalogue, so two majors. Two majors sit between the checklist's one major (10) and several (0), so 5, as deepseek-api, black-forest-labs and telnyx-voice are read (5). The 19 October check gave the same 5 because no 90-day history could be read. A 30 October incident on deployment creation, dedicated training and batch inference was still open on 1 November. Numeric adaptive ceilings by model size, 216M total prompt, 54M uncached prompt and 2.16M generated tokens a minute for Small models (15). The rate-limits page says to honour Retry-After on 429 and otherwise back off with jitter, and the reliability guide covers timeouts and retries (15). Reserved throughput is 'covered by a throughput SLA' with no figure published (5 of 10). Serverless is generally available (10). 20+5+15+15+5+10 = 70.

Amazon Bedrock 30

No x402, MPP or L402 (0). Per-token prices per model, tier and Region on the public pricing page, filled from AWS's public price feed (20). New accounts get up to $200 in Free Tier credits and the Free Tier page lists Bedrock among eligible services, and the Free Tier FAQ says most new customers don't need a payment method, but AWS may still ask for one, so half, as for S3 (10). A person signs up for AWS in a browser and sets up IAM before any key exists (0).

Fireworks AI Inference 40

No machine payment protocol in the docs index or billing pages, re-read on 1 November (0). Public per-token prices by model and serving mode (20). The credits FAQ documents a $1 starter credit that runs without a payment method, after which the account is suspended until one is added (20). Account signup and credit purchases need a person (0).