Titan vs Claude on Amazon Bedrock: Wetin You Go Pick?

Wetin Titan really be on Amazon Bedrock today (September 2026)

If you dey Amazon Bedrock and you dey ask "Titan vs Claude, which one I go use?", the first thing you need hear na this: as of September 2026, Amazon Titan text generation models no dey the Bedrock model catalog again. Wetin remain for Titan side na Titan Text Embeddings V2, Titan Embeddings G1 Text, Titan Multimodal Embeddings G1 and Titan Image Generator G1 v2. Titan Text Express, Titan Text Lite and Titan Text Premier no appear for the "Models at a glance" page at all. Amazon first-party text model now na Nova, and the newest one for that line na Nova 2 Lite.

So the real question wey person dey ask when e talk "Titan vs Claude" na two questions join together. One: for text generation, na Nova vs Claude. Two: for embeddings (the number vectors wey you use for search and RAG), na Titan Text Embeddings V2, because that one still dey active and na wetin most people mean when dem say "Titan" today. Make we answer both, then land on the pairing wey most Bedrock RAG builds dey use: Titan for the vectors, Claude for the answer.

Check am yourself before you plan anything, because model status dey change and the API go show you the truth for your own account and region:

aws bedrock list-foundation-models --by-provider Amazon \
  --query 'modelSummaries[].[modelId,modelLifecycle.status]' --output table

The modelLifecycle.status column go show ACTIVE or LEGACY for each model ID. LEGACY mean the model don get EOL (end of life) date, new customers no fit start am, and nobody fit create new provisioned throughput for am. Bedrock keep the full legacy list on the model lifecycle page, and every model card carry one "EOL no sooner than" date. For example, Claude Sonnet 4 enter Legacy on April 14, 2026 with EOL of October 14, 2026, so anybody wey still dey call am need move to Sonnet 4.6 or Sonnet 5. Read that date before you build anything on top a model, to avoid wahala later.

Which capability tier you dey compare: Opus, Sonnet, Haiku vs Nova

Claude on Bedrock get three tiers wey you go see everywhere. Opus na the biggest brain. Sonnet na the everyday worker. Haiku na the fast and cheap one. As of September 2026, the Bedrock catalog list Claude Opus 5, Claude Sonnet 5 and Claude Haiku 4.5, plus the Claude 4.x line (Opus 4.8, Sonnet 4.6 and others) and Claude Fable 5.1 on top of Opus. Bedrock own model card describe Opus 5 as the model wey dey power long-running agents, and Sonnet 5 as "near-Opus intelligence" for coding and professional work at scale. If you no sure which tier fit your job, our guide on when to pick Opus, Sonnet or Haiku break the decision down.

Amazon side: first-generation Nova get Micro (text only, lowest latency), Lite, Pro and Premier. Then Nova 2 Lite land for December 2025 with extended thinking and better tool use, and AWS position am as "cost-efficient multimodal model for simple automation, document processing, and customer support". Nova Premier v1 dey the Bedrock legacy table with EOL date of September 14, 2026, so no start any new build on am.

Benchmark scores: no number go appear here, because the only honest benchmark na the one you run on your own prompts and your own documents. Wetin the model cards give us na positioning. Nova 2 Lite na Amazon budget tier. Claude Haiku 4.5 na Anthropic budget tier, and e dey the same catalog with Sonnet 5 and Opus 5 wey sit on top am when the job get harder. If your task na to summarise support tickets or classify text, test Haiku 4.5 and Nova 2 Lite side by side with fifty real examples. If your task na agent wey go run tools for many steps, or serious code work, the model cards themselves point you to Sonnet 5 or Opus 5.

How much each one cost per 1M tokens on Bedrock

Money na the thing wey decide most builds, so make we put the figures for one place. Bedrock bill Claude through AWS Marketplace, so the charge go appear on your AWS bill under Anthropic, not under Amazon Bedrock. Nova and Titan charges appear under Bedrock itself.

[
  {
    "label": "Claude Opus 5",
    "input_usd": 5,
    "output_usd": 25,
    "notes": "Anthropic list price, Sept 2026"
  },
  {
    "label": "Claude Sonnet 5",
    "input_usd": 2,
    "output_usd": 10,
    "notes": "Anthropic list price, Sept 2026"
  },
  {
    "label": "Claude Haiku 4.5",
    "input_usd": 1,
    "output_usd": 5,
    "notes": "Anthropic list price, Sept 2026"
  },
  {
    "label": "Nova 2 Lite",
    "input_usd": "0.30",
    "output_usd": "2.50",
    "notes": "Public price tracker figure, Sept 2026"
  }
]

Claude Opus 5 na 5 dollars per 1M input tokens and 25 dollars per 1M output tokens. Sonnet 5 na 2 dollars in and 10 dollars out. Haiku 4.5 na 1 dollar in and 5 dollars out. Nova 2 Lite published figure na 0.30 dollars in and 2.50 dollars out. The Claude figures na Anthropic own list price as of September 2026, and Anthropic pricing page note say the Sonnet 5 introductory price don become the standard price. The Nova 2 Lite figure come from public price trackers on the same date, because the Bedrock pricing page load the tables with script and we no fit quote am line by line here. Before you budget, open the Bedrock pricing page for your own region and confirm every number.

Two things about Bedrock pricing wey people miss. First, Anthropic pricing docs say Bedrock regional endpoints (the us. and eu. inference profiles) carry 10 percent premium on top the global endpoint price for Claude 4.5 and later. So global.anthropic.claude-sonnet-5 na the cheaper route if you no get data residency rule. Second, output tokens cost five times input tokens for every Claude tier in the chart, so the shape of your prompt and the length of the answer matter more than the model name. How input and output tokens hit your Claude bill differently show that with worked examples.

For the embeddings side, Titan Text Embeddings V2 price na small change. AWS announce am for May 2024 at 0.02 dollars per 1M input tokens, which na 0.00002 dollars per 1,000 tokens. That mean you fit embed one million tokens of your documents for two cents. Confirm the current figure on the Bedrock pricing page, but this na why the Titan plus Claude pairing make sense: the cheap part handle the volume, the expensive part handle only the final answer.

If you dey also weigh OpenAI before you commit to Bedrock, Claude vs ChatGPT API pricing carry the same per-token view for both sides.

Context window and tool use: wetin fit enter and wetin the model fit do

[
  {
    "label": "Claude Opus 5",
    "context_tokens": "1,000,000",
    "max_output_tokens": "128,000"
  },
  {
    "label": "Claude Sonnet 5",
    "context_tokens": "1,000,000",
    "max_output_tokens": "128,000"
  },
  {
    "label": "Claude Haiku 4.5",
    "context_tokens": "200,000",
    "max_output_tokens": "64,000"
  },
  {
    "label": "Nova 2 Lite",
    "context_tokens": "1,000,000",
    "max_output_tokens": "64,000"
  },
  {
    "label": "Nova Micro",
    "context_tokens": "128,000",
    "max_output_tokens": "10,000"
  }
]

Claude Opus 5 and Sonnet 5 both carry 1,000,000 tokens of context window and 128,000 max output tokens on Bedrock. Haiku 4.5 na 200,000 context and 64,000 output. Nova 2 Lite match Opus and Sonnet on context but stop at 64,000 output. Nova Micro, the small first-generation model, get 128,000 context. For RAG, big context mean you fit pass more retrieved chunks inside one call, but e no mean you should. Every chunk you pass na input token wey you dey pay for.

Tool use (the model calling functions wey you define, so e fit fetch data or take action) dey supported on both sides through the Bedrock Converse API. The Claude model cards list Agents, Knowledge base, computer use and prompt caching as supported on bedrock-runtime. The Nova 2 Lite model card list client-side tool calling and prompt caching as supported, but e mark Bedrock Knowledge Bases and Bedrock Agents as not supported as of September 2026. That one matter: if your plan na to use Bedrock managed Knowledge Bases for RAG, the generation model for that flow need to be one wey Knowledge Bases support, and Claude Sonnet 5, Opus 5 and Haiku 4.5 all pass that check.

Prompt caching also differ. Claude Opus 5 minimum cache checkpoint na 512 tokens, Sonnet 5 na 1,024, Haiku 4.5 na 4,096, and Claude support both 5 minute and 1 hour TTL (time to live). Nova 2 Lite cache na 5 minute only, with 20K token cap. If your RAG system prompt big and e repeat for every call, caching go cut your input bill well well, and the prompt caching break-even guide show you the point where e start to pay back.

Regions: where you fit call each model from

This na the part wey Nigerian builders need read twice. Bedrock get three ways to call a model. In-Region keep your request inside one AWS region. Geo cross-region (the us., eu., au. and jp. prefixes) route inside one geography. Global cross-region (the global. prefix) route anywhere wey capacity dey.

Claude Sonnet 5 and Haiku 4.5 on bedrock-runtime require geo or global inference profile ID. The model card talk am plain: "The bare model ID isn't supported for on-demand throughput." So you go call global.anthropic.claude-sonnet-5 or us.anthropic.claude-sonnet-5, never anthropic.claude-sonnet-5. The global profile source regions include regions in Africa, the Middle East, Europe and Asia, so you get many entry points.

Titan Text Embeddings V2 na the opposite. E dey in-region only: the model card mark geo and global inference as "Not supported". You must pick one region wey list am and send every embedding call there. So a RAG build on Bedrock normally run two clients: one bedrock-runtime client pinned to one region for Titan, and one client wey use a global Claude profile. Both fit be the same boto3 client if the region you pick for Titan also appear in the Claude global source list, and most of them do.

Nobody go tell you here which region na the nearest for you, because latency depend on your ISP route and AWS dey add regions and models every few months. Open the "Regional availability by models" page on the Bedrock docs, find amazon.titan-embed-text-v2:0, and note the regions wey carry am. Then confirm the same region appear in the Claude global source list on the Claude model card. Ping both endpoints from your VPS and go with the numbers you see, not with wetin anybody post on social media.

On-demand vs provisioned throughput: wetin still dey available

Bedrock don move from the old "on-demand vs provisioned throughput" split to service tiers. Standard na pay per token, no commitment. Priority na faster response for a premium. Flex na cheaper price for jobs wey no dey rush. Reserved na dedicated throughput with a term commitment, set at account level by talking to your AWS account team. The old Provisioned Throughput still exist for some models, but you no fit create new one for any model wey already enter Legacy.

As of September 2026, the model cards say: Claude Opus 5 and Sonnet 5 support Standard only, and Opus 5 also list Batch. Claude Haiku 4.5 support Standard and Reserved. Nova 2 Lite support Standard, Priority and Flex. Titan Text Embeddings V2 support Standard. So if you want Flex discount for background jobs, Nova 2 Lite get am and the Claude 5 models no get am on Bedrock today. If you want committed capacity for a Claude model, Haiku 4.5 na the one wey list Reserved. For most small teams, Standard on a global Claude profile plus Standard Titan embeddings na the whole story, and you fit skip this section until your bill pass a few hundred dollars a month.

Titan vs Claude for RAG: the pairing wey most Bedrock builds use

RAG (retrieval-augmented generation) na the pattern where you turn your documents into vectors, store them, find the ones closest to a question, then give those chunks to a language model to write the answer. On Bedrock, the common pairing na Titan Text Embeddings V2 for the vectors and Claude for the writing. E common because Titan V2 cheap, e accept up to 8,192 tokens per input, e output 1,024 dimensions by default with 512 and 256 as options to save vector database space, and e carry over 100 languages in preview, Hausa and Yoruba included. Claude then read the retrieved chunks and answer, with tool use and long context when you need them.

Here na the minimum wey work. Install the SDK:

pip install boto3

Authenticate. The quickest way for testing na a long-term Bedrock API key from the Bedrock console, exported as AWS_BEARER_TOKEN_BEDROCK. For production, use an IAM role or aws configure credentials with a policy wey allow bedrock:InvokeModel. If you dey mix Bedrock with the first-party Claude API, how Claude API authentication works cover that side.

export AWS_BEARER_TOKEN_BEDROCK="<your Bedrock API key>"

Embed one chunk with Titan:

import json
import boto3

# Pick this region from the Bedrock region table. Titan Embeddings V2 na in-region only.
embed = boto3.client('bedrock-runtime', region_name='us-east-1')

resp = embed.invoke_model(
    modelId='amazon.titan-embed-text-v2:0',
    body=json.dumps({
        'inputText': 'Our refund window na 14 days from delivery.',
        'dimensions': 1024,
        'normalize': True
    })
)
body = json.loads(resp['body'].read())
print(len(body['embedding']), body['inputTextTokenCount'])

You should see 1024 and a small token count print out. If you see ValidationException about the model ID, the region you pick no carry Titan V2, or your account never enable model access for am in that region.

Now answer with Claude, passing the retrieved text as context:

import boto3

gen = boto3.client('bedrock-runtime', region_name='us-east-1')

context = 'Our refund window na 14 days from delivery.'
question = 'How many days I get to return a product?'

resp = gen.converse(
    modelId='global.anthropic.claude-sonnet-5',
    messages=[{
        'role': 'user',
        'content': [{'text': f'Use only this context to answer.\n\nContext:\n{context}\n\nQuestion: {question}'}]
    }],
    inferenceConfig={'maxTokens': 512}
)
print(resp['output']['message']['content'][0]['text'])

A healthy result na one or two lines wey quote the 14 days. If you get AccessDeniedException, your account never accept the Anthropic model terms in the Bedrock console, or your IAM policy no cover the inference profile. If you get ValidationException wey say the model no dey supported for on-demand throughput, you pass the bare anthropic.claude-sonnet-5 instead of the global. profile.

The vector store and the retrieval loop between those two calls na the real work, and e no depend on Bedrock at all. Our guide to self-hosting a RAG pipeline on a VPS build that middle part with a vector database wey you own, and you fit swap the embedding call for Titan and the generation call for Claude exactly as above. If you want Amazon to manage the middle for you, Bedrock Knowledge Bases do the same job, and na there the Nova 2 Lite "Knowledge base: not supported" line become important for your model choice.

One more option wey enter the catalog: Amazon Nova Multimodal Embeddings, wey embed text, image, document, video and audio into one space. If your documents na mostly scanned PDFs and pictures, test am against Titan V2. For plain text, Titan V2 remain the cheap and proven default.

Payment and account gist for Nigerian builders

AWS bill in US dollars. The Claude charges appear as AWS Marketplace line under Anthropic, the Nova and Titan charges appear under Bedrock, and everything settle on the card or payment method attached to your AWS account. Nobody go tell you here which Nigerian bank card go pass or fail, because that policy dey change bank by bank and month by month. Wetin fit help na wetin to do when a payment bounce: our guide on what to do when a Claude payment dey declined cover the checks (card currency and international transaction limits) wey apply the same way to AWS.

Set one alarm in AWS Budgets before you run your first embedding job. A full re-index of a big document set na the kind of thing wey surprise people, because output tokens on the generation side cost far more than the embedding side, and a retry loop wey you forget go run all night.

If your team already use Claude Code for development, you fit point am at Bedrock with CLAUDE_CODE_USE_BEDROCK=1, so the coding agent and your RAG app share one AWS bill and one set of IAM roles. Running Claude Code through Bedrock or Vertex walk through that setup.

FAQ

Titan Text still dey available on Amazon Bedrock?

As of September 2026, no. The Bedrock "Models at a glance" catalog list Amazon Titan Embeddings G1 Text, Titan Text Embeddings V2, Titan Multimodal Embeddings G1 and Titan Image Generator G1 v2 only. Titan Text Express, Lite and Premier no appear there. Amazon first-party text generation line na Nova, with Nova 2 Lite as the newest model. Run aws bedrock list-foundation-models --by-provider Amazon to see wetin your own account fit call in your region.

Which one cheaper for RAG, Nova 2 Lite or Claude Haiku 4.5?

On published figures for September 2026, Nova 2 Lite input and output tokens cost less than Claude Haiku 4.5. But the RAG bill dey split in two: embeddings (Titan V2 at about 0.02 dollars per 1M tokens) and generation. Generation output tokens dominate, so test answer quality on your own documents first. A cheaper model wey make you re-ask twice no cheap. Also check the Nova 2 Lite model card: Bedrock Knowledge Bases dey listed as not supported for am, while Haiku 4.5 support am.

Why my Claude call fail with "model ID isn't supported for on-demand throughput"?

Because you pass the bare model ID, like anthropic.claude-sonnet-5, to bedrock-runtime. The Claude Sonnet 5 and Haiku 4.5 model cards require a geo or global inference profile ID for on-demand calls. Use global.anthropic.claude-sonnet-5, or us.anthropic.claude-sonnet-5 if you need US data residency, and remember the regional profile carry a 10 percent price premium.

Titan Text Embeddings V2 dey work with cross-region inference?

No. The Titan Text Embeddings V2 model card mark geo and global inference profiles as "Not supported". You must call am in-region with amazon.titan-embed-text-v2:0 and a region_name wey appear in its regional availability table. Pick that region from the Bedrock region table, then confirm your Claude global profile also list that region as a source, so one boto3 client fit serve both calls.